Projects
![]()
No projects found
Investors
New
DAO Treasuries
DATs
Lend / Borrow APRs
Chains
Categories
Spaces
Decentralized AI Agents with Oasis, Omo, & Giza
Recorded: July 30, 2024 Duration: 0:38:57Space Recording
Full Transcription
Hello.
Let's just give it a minute for people to join and then we can kick off.
We're still missing Fran from Giza, but he'll join in a sec.
Since I'm speaking for Oasis, generally this means it's an AI topic.
This time we're going to talk about decentralized AI agents.
Last time we talked quite a bit about the general space, like how does crypto and AI make sense?
What are the problems in centralized AI?
Why is decentralized AI better?
Or rather, what kind of problems does it solve?
And everyone talks about GPUs.
Everyone talks about user-owned AI.
And I think the third big topic that people really kind of get caught on is on-chain agents that are powered by off-chain AI.
And there we brought some experts that are focusing solely on this.
Giza and Omo.
As I said, Fran, the founder of Giza, has to join still.
And I'm really excited to make this a quite educational session.
I have lots of questions in the space still because we mainly build kind of infrastructure so that people can build on top.
Everyone knows, I think, here at Oasis, we have our own confidential EVM that unlocks quite some unique AI use cases.
So maybe let's start with you, Ali.
Do you want to give us an intro about yourself and what you guys do with Omo?
Yeah, for sure.
GM, guys.
I'm Ali Erard, whichever is easier.
I'm co-founder and CEO of Omo.
Omo Protocol is an agentic execution platform.
So what this means is these agents are essentially EOAs, externally-owned accounts.
They can own their own assets or control users' assets in this case.
And then they get instructions off-chain from an ML model or a static Python script within the cloud server.
So that's kind of about on-chain agent automation, essentially, in DeFi.
And why is it specifically DeFi?
DeFi is the lowest-hanging fruit.
We are exploring automating governance.
I mean, I'm sure you guys just saw the compound hack as well.
If you had agents to automate your governance decisions, you can protect attack vectors like that.
But yeah, DeFi is the lowest-hanging fruit.
Obviously, we want the space to expand into more sectors and niches like GameFi, GammaFi, etc.
But DeFi is just the lowest-hanging fruit and kind of proves the proof of concept.
Right.
I mean, you're one of the few projects that is actually building something.
Like, most projects are talking about AI.
They're building infra.
And we're all just waiting for projects to build.
So I'm sure Fran, that just joined, and we as well from Oasis, are really happy to see someone actually building something.
Yeah, no, Fran is great.
Definitely not possible without him.
He's been a super great help to the Omo team in general.
So super happy to work with, you know, the Giza team and the Oasis team, respectively.
Happy to hear.
Then, Fran, once we get you up here as a speaker, feel free to also intro yourself and Giza, what you guys do, what you focus on.
Yes.
Hello, guys.
Hello, everyone.
Thank you for having me here.
I'm a pleasure to chat with all this collaboration that we are doing and enjoying all about Giza.
So I'm Fran.
I'm one of the co-founders, mostly in charge of all the technical part that we do at Giza.
My background is on AI.
I've been doing it for 10 years before it was cool.
And now in the crypto space, doing agents.
That's what we are, like, creating at Giza.
But not, like, typical agents.
The main difference about agents is that they are, like, verifiable, which means that they are, like, static, deterministic.
And every decision that they create for interacting with DeFi protocols, it's completely verifiable.
So every interaction, every action, for example, with our friends at Omo, is going to generate a secret proof in order to have more transparency and the station of how the agent is operating,
which is something that currently no one has.
Great.
I mean, I'm sure we're going to talk about the verifiability later on as well, since that's obviously very close to my heart and everyone at Oasis, because it's what we think about all day.
But I mentioned in the intro where you were missing that I really want to make this an educational one.
So maybe we can start with the basics.
Like, if the AI stack is from bottom to top, like decentralized compute, then we have foundational models, then we have custom data or private data that you can use to then for inference results or to build fine-tuned models.
And at the top, we have applications.
How does this look like in the agent space specifically?
Does it differentiate somewhere?
Maybe frown you.
Yes, we take it like a different approach to agents, where we don't take LLMs as like the entry point of interacting with, in this case, with financial products, because in the end, in this space, you need like availability of what the agent is doing.
And then LLMs and LLMs for now don't provide this kind of environment, especially if you are going to ape into a DeFi protocol.
So, and our basic stack ends when we have all the on-chain data in order to create like an action, then we have like a computation layer that can be using a simple machine learning model that is reliable, but also combined with heuristics in order to make a strategy or create some kind of interaction with multiple DeFi protocols.
But it's not something that you need to prompt into a chain.
And would you consider these heuristics intelligence?
Like how, what are we talking about?
In the end, it's about like creating like data-driven decisions.
So based on something that is happening on chain, they create some kind of action, like for example, like depositing in Trinomobolt or swapping in JuniSwap or lending on an Aave market.
so here this heuristic defines like what are the conditions for which action to happen
based on what is happening from on-chain data but also from any kind of off-chain
kind of market situation or data that the agent needs to take into account
understood okay maybe for you ali since you're building agents is the infrastructure specifically
for the asian economy that we currently have provided by for example giza enough to do what
you want to do or did you have to take some kind of shortcuts or some stuff is somewhere down the
roadmap just because the stack isn't ready yet no no giza provides like a pretty valuable service i
mean we've been working in crypto ai since the start of the year and we've been messing with
different agent stacks agent providers um and you know it's been kind of difficult you know there
there there are multiple agent stacks with their own constraints who just found that the you know
the giza team has been the most adaptable and flexible to work with us so we can actually
you know bring agentic applications to market because you know this stage of like the product
i'd say is the most difficult going zero to one and actually implementing you know static or dynamic
agents into a defy product uh is difficult because we don't have a framework or like you know workflow
of how we can do it consistently once we get the mvp done and we can replicate this process
then it becomes much easier it's really important to have like a really good relationship with the
agent stack team respectively so no giza can definitely fulfill the role um i think like the
one area where giza might be lacking is primarily uh with the verification of inference if we want to use
like these larger models like llama or mistral um then it's going to be like okay well it's not
realistic to do zkml for that we'd have to look into like tml potentially op mobile opml isn't the best
right i mean it wasn't supposed to be a shot at giza at all i was just curious in general like when
you started looking for partners here if something was missing and that's kind of uh where you adjusted
and pivoted yeah no for for the most part i mean the uh you know working with the team and you know uh
the static agent implementation should be quite effective i mean it's you know it's really important
to take like a stepwise manner to actually implementing machine learning models into like
actual crypto applications especially handling you know uses our customers funds so that's why i like
the static agent implementation is going to be uh quite prolific for us where you know we can kind of
go to market have like a deterministic agent and then as we see this vault product market fit uh we
can have a more quote unquote risky or volatile vault where we use a machine learning model on the back end
via you know giza's dynamic agent implementation uh and obviously you know we'll have you know um
don't know the exact word uh essentially cautionary details are like okay this is using machine
learning this is not a deterministic model you know this is machine learning you know it's going
to be essentially random based off of the outputs or the you know the outputs of the calculations it makes
so yeah right and maybe you you guys can lead us through the process of someone actually building
on top of an agent info provider like what do they need to bring and what's already there and then how
does the distribution happen towards like different communities yeah no for sure working with
the agent stack providers it's going to be really different um on a sdk or cli basis so um for the
context of giza um in the current cli you know you have to you know host the model in your own local
machine and then you know then you can be able to uh you know deploy an inference endpoint via the giza
cli and then once the endpoint is prepared the endpoint is used to basically verify uh you know the
inference output and then also used to basically uh submit these proofs on chain uh respectively um
so you know the going back to the first question it's quite complex and it varies from agent stack
to agent stack provider i mean we're speaking with the tallest network team as well but these guys uh
i think are like an internal devnet so we haven't even worked with a stack at all um
to be frank uh you know the giza guys are the furthest along and like the proper you know agentic uh
crypto ai stack that that we see in the space so far so there hasn't really been a comparison uh
honestly
okay i mean you guys focused on d5 for a reason but which kind of agents i mean it could be d5 as
well maybe some sub segment maybe outside which agents make sense in a decentralized setup in an
on-chain setup and which ones actually don't i guess could you expand further on like the decentralized
portion
sorry was a question to me uh yeah yeah it was to you marco
yeah i mean why built it on like a decentralized compute layer plus an agent infra for something to
interact with different blockchains instead of i don't know building this in a centralized off-chain way
yeah i know that's a good question so i guess um when we talk about decentralization and crypto ai
uh you know sure using you know decentralized compute layer is good um i i think like you know it really
goes back to like these agents primary use case it's just for better ux i i don't really think um
you know decentralization is that big of a priority because inherently how the omo protocol works is if
a user deposits into omavolt you're gonna have an agent withdraw that capital on behalf of the user and
then execute their respective field strategy so this is to a certain degree custodial so this isn't
necessarily decentralized obviously we're looking to uh you know leveraging um you know oasis's t uh to
mitigate the amount of quote-unquote custodial risk there is uh but you know on paper it you know
to certain people it can look custodial so um i don't know decentralization is the right uh thing
here it's more so leveraging agents can uh you know abstract away a lot of the main issues with
crypto say outside legalities ux um when we think about like how grandmas or grandpas can use apple
products with simplicity and you know there are like nine to twelve different steps a user needs to
execute even if they're a power user if they want to do some complex strategy uh whether it's like you
know hedging a lp position of like a token like usde uh in case of like a deep x scenario right if you
want to unwind that position like it's going to take you like 15 to 20 minutes to completely get this
position set up and then unwinding the position is like another 15 to 20 minutes so it goes back to
how agents the primary use case is for just better ux in the industry
fran what do you see built upon giza like what kind of use cases are there
and i totally agree with you ali like ux is definitely the one where we can improve the most
uh yeah also to span on the decentralization part uh that decentralization also can affect
latency of the solutions that you build so you also need to be like uh taking consideration what kind
of solution you're building or use case you're building um so the approach that we are following
is instead of creating like this decentralization of compute or using a decentralized compute layer
we just use ck for the whole like agent so everything is as trustless as possible by providing these
safety proves and making them available for anyone to verify that the agent executed correctly and there's
like no trash assumptions or even like latency overhead like this executing in a decentralized environment
in terms of uh use cases uh most of the demand or with of the protocols that we're working with is
how they can provide one of their ux for the users by creating uh automations on top of their protocol
or by providing like an abstracted way to interact with uh these kind of protocols or a way to avoid
smart contract upgrades uh into these protocols by providing agents that do this specific logic in a
trustless way by executing in a verifiable manner that is deterministic so basically uh outsourcing what
is usually a smart contract to a ck application that is run in the agent infrastructure that we have
um helping these protocols be like more uh iterative by providing better functionality and providing better ux for
for the users and users here are developers on different protocols
yes we have like a different kind of use cases most of them are like regarding defy
protocols or directly with them you know to create these new experiences or dux
for uh for them but also it's very oriented to uh to the consumer side of things because one of the
uh things that uh we're going to be releasing very soon as well for organs is that you don't need to
have uh this delegated kind of uh environment where you send your funds to an agent and then it's like
it just acts now the agent will be able to just you connect your wallet and the agent can automate
uh inside your wallet directly without uh withdrawing any kind of funds uh in automations on top of any
kind of defy protocol so instead of uh the clear example would be uh i'm on a chain or an ecosystem where
there is no uh options protocol or gel aggregator so instead of that i can leverage the usage of
landing markets and amms to create the same functionality as a ck unverifiable agent that runs
off chain but with ckplus being verified on chain and automates any kind of strategy or functionality
that they use for once directly from his wallet okay so it makes sense to think of agents just kind
of automation tools for us to use in the crypto world and then there will be an automation marketplace
like zapier what we have and web2 where you can connect different um yeah automations that you need
but from the consumer perspective crypto in general is kind of scary like even i get try to get well i
almost get scammed like on a weekly basis because i think i got an airdrop but then it was something like
fake how can i trust these agents like when will we get to a point where i can see like
hey this is an agent that will get me the best yield for my stablecoin that i'm paid with from oasis
and that i just want to use but i actually don't it's a headache for me to find this manually so i i
this is something where i would like to use an agent but how do we get to this level of like hey
you can here use an agent you don't need to delegate but they will actually work even with your wallet
but make this like trustworthy it's something i'm really struggling with
that's a very good yeah uh this is a very good point actually uh marco that's for example one of the
issues that i see with any kind of system that relies on on llms to doing launching transactions and is that
you you may ask the uh the island to okay i want to swap my usdc into it and then bring it to
a base and then like put into a bit but maybe if there is something and grown in the middle like
an elucination of the lm it may do whatever thing with your money and you don't know really what what
is going on and the the bad thing is that once that happens that transactions happen you need to
manually like uh unwind what the agent did so in the end there is no actual benefit for for these
solutions and in the end they when the user prompts to the lm needs to really know uh what the user wants
and most of the cases of why a user like that will use an lm is is because of the complete opposite is
because the user doesn't know what to do in terms of interacting with d5 protocols so um
so yeah coming back to uh how to make this solution trustworthy uh in our in our case uh i think that
providing uh very good um understanding of how the whole underlying like automation strategy or
or basically that had driven decision that the agent is taking uh needs to be transparent
and reliable so that's why its execution is threshold by a city proof and also linked to the
transactions that the agent is making so it's not that you need to trust us but there is like evidence
and proofs that that happens too yeah just to expand upon that a bit more as well um you know to make
you know these machine learning models more robust and more verifiable it's just we need to test them
uh you know we just need to go through the iterative feedback loop of seeing if models do
hallucinate or you know provide poor outcomes um what went wrong how can we improve it uh that's just
the reality of the crypto ai sector it's extremely nascent the agent space is even more nascent um
and we just need to uh test these products and like let our users know that there may be risks associated with it
um we are also speaking to uh our service provider symbiotic i can't dive too deeply onto the details of
this but you know looking to see how we can leverage uh you know slashable uh uh slashable economic
security framework attached to you know these agents that might be you know holding a lot of tvl um ensuring
that these these uh dynamic uh votes uh match apr or exceed the static agent implementation and if
they do then that could trigger a splashing event because you know in theory it'd be assumed to be
that these machine learning model votes should at least perform on par if not better than these static
agent implementations so that's another thing we're exploring uh to potentially reduce um you know poor
outcomes for these machine learning models there's also rag i know a couple companies that are working
on rag within the crypto ai space but that's also extremely nascent industry too there there are more
things coming out to help mitigate the risks associated with using these models but you know we're all
early we're all early in crypto we all know the risks associated with crypto it's just basically the
same exact paradigm except that there's just some risks using these crypto ai applications
got it i mean ai explainability is bad anyways ai alignment is still not something that's solved
even remotely so i totally understand that we actually prefer i assume um and kind of correct
me if i'm wrong we prefer agents with slightly less intelligence like they don't need to be powered
by llms they just need to do what we actually want them to and for that kind of hard-coded models
uh or very simple small models uh are enough okay interesting because and you guys mentioned this a
couple of times like the verifiability of all of this that happens off chain is still lacking like
inference marketplaces today we just create these jobs on chain we send the compute job off chain
some nodes do the compute and they send us a result back but it's still a black box to most
users and kind of projects as well because we don't know if they actually use the data that we
send them to compute on a model that i expect them to use and the result that they're sending me
is actually the one that i expect or is the one that was computed so fran you mentioned a couple of times
like zk for this output verifiability did you look at other stuff does it actually solve the big problem
that we have or is it just like a current fix for like hard-coded stuff for simple models
that's a that's a really good point um the the problem with using uh this kind of computation for
for aliens is that uh most of the things may need to be like public or the data that you send for
computation or the model itself to be available to multiple nodes so this open apps the window for
also the inference run from your agent so if you're like in a use case where like you're just trading
options uh you just want like super low latency and if you are like using a this compute network
you might not have the right performance requirements you know to run this kind of use case
but also neither the privacy guarantees as you said marco uh so um our stance is that we use ck because it
can provide us like a better performance even the complexity of the uh of the strategies that they are
running but also uh the right privacy because what the the agent is doing in terms of the logic of the
strategy can be public but also can be private so it's the users that indian are using the agent
are not like front run by no one trying to do mb or any kind of other thing in general
got it and maybe just a sneak peek for everybody into kind of what we're working on in oasis so we have
our own l1 we have a confidential evm our whole kind of logic is built on te's trusted execution
environments that have recently gotten lots of the voice of share and a positive one which is very
surprising to us because normally people just hate on te's because there is some hardware dependency and
people just don't like having a dependency on for example intel which is our case where
the nodes have intel sgx cpus but yeah people tend to ignore then if you have different defense in
depth mechanisms in place but what we are looking forward to or working on is kind of expanding the
te's into also including gpus so you can have a cpu plus gpu te and with this you can have like
proper llm compute in a trusted execution environment and fully verify that the data was used correctly it was
transferred in a confidential manner and the results are like end-to-end verifiable and this
actually like works really well with crypto like normally everything in crypto is verifiable and
decentralized it doesn't mingle too well with non-deterministic things and that's currently
what's happening a lot with this intersection of crypto xai so this just kind of as a sneak peek
um keyword user onboarding i mean this is what everyone is hoping for like hey
how do we get web2 users to use crypto is the agent economy something that's going to help
or are all of these automations just built for web3 users
i can take this um and i'd like to hear your opinion as well fran but yeah i think uh you know
depends on the product of course like obviously something as niche as like you know yield farming
using d5 primitives or on-chain primitives on like bear chain or monad that's obviously a lot more
tailored to web3 audience but like um you know meme coins you know you can see it in like you know
south korea and also the us right they have quite a tangible appetite people that aren't that crypto
savvy or trading meme coins celebrity coins as well so it depends on the use case for sure and automation
is just going to agentic automation is just going to result in better ux um in general uh so yeah it
could you know more products need to get built out outside of just like the d5 meme coin trading
paradigm like you know we need more social fi we need more uh more prediction markets more gambling
honestly uh for you know more web2 users to get on board and then agents are just going to help
uh seamlessly transition users to using web3 applications without them knowing like which
rpc they're using uh what wallet they need to use uh what network they're on etc etc
from my side i think uh what agents enable is solving the problem of discovery in web3 when as a
newcomer that just uh dreams crypto and you go into the whole ecosystem you have like hundreds
hundreds of lending protocols dozens of amms so what do you trust what do you use you don't even know
right so i think agents can really abstract this kind of layer where you just provide the right
strategies to have like a better onboarding that is super close to the wallet level uh where just
users come one click and you just deposit your uh you just interact your if for the with the agent and
the agent just interact with whatever protocol that is trusted and and reliable by the agent developer
right and the other kind of uh problem that this solves um is really onboarding and i think this is
super good in terms of using agents for example in in chains like base uh where and the problem comes to
kyc if if you want to start like using if or by your first if you need to like be
doing a kyc process to buy on ramp and and for most of the people this is like a huge barrier
and now like in base if you just have like a coinbase kind of account you can you don't need to have any
kind of kyc additional kyc in order to interact with the base and any kind of protocol there and if you
combine that with usdc free payments and agents i think there's like a huge opportunity there in
terms of solving the ux part for um for crypto
it actually reminds me of something that i saw not too long ago do we actually need payment rails
between agents for them to interact autonomously with each other
i think that's the blockchain to me the blockchain is the payments rails of
agents so i think it's like the perfect environment for these for agents to interact with payment systems
right i mean obviously i i was just seeing projects building specific
let's say blockchains roll ups that are custom built for like kind of agent interactions which
is likely going to be like very small amounts very quick you need finality rather quickly so okay
but if you say current infra is enough then i'm very happy to hear because i don't know if we need
that many more blockchains
i think agents needs to be uh one where the liquidity is in order to create the best experiences
but also where the users are and creating like uh but that's like a personal new rollups maybe creates
more fragmentation in terms of liquidity and users all this we explore more like this concept of
chain of structure and and all the potential benefits that uh they have um it's going to be harder like
to create yes like rollups for agent payments or similar things i think we just need the right
infrastructure to interact with the existing uh world chains okay totally agree um what would you guys
like to see built in this agent space do we need more infra do we need more agents do we need an agent
marketplace do we need just like very simple interfaces for users to pick agents uh what's
missing like maybe bridging isn't good enough because bridging is always a risk factor currently
and people like try to avoid it and stay in one ecosystem but i guess for agents to be really
effective you need really good bridging to different ecosystems to you kind of get the most out of it
uh that's a good question i guess so one one thing to address the bridging issues we're speaking with
shogun fight these guys are an intent aggregation platform uh so like their solvers uh are quite
robust like you know with centralized exchanges so that are you know our implementation of agents are
going to be leveraging uh shogun finds as if you know traditional arbitrary messaging uh passing bridges so
uh we'll see how that performance is but i think that could result in not only better ux but also
enhanced security and lower slippage for the end user uh thinking about like what kind of agents like
what's missing in like the agent spaces um it's a good question um i don't know i i just honestly
want more people to build like agent stacks that have like their own you know private and public key
uh you know the more the better uh you know the more the better i'd be curious to see what other
teams like kind of iterate upon you know existing agent stacks uh if they have a different take of
how to execute upon it to me what's missing is applications i think everyone is trying to build
infrastructure layer one for agents or whatever for agents and if everyone builds infrastructure but no
one's use it or everyone is targeting developers but there are no developers no one is going to use
anything so we don't need to focus on creating applications that leverage the potential of agents
to create like this better ux and i feel like combining working with these amazing people of homo and also
what using the infrastructure of wasis can really enable this new set of applications for agents
yes building applications that's what everyone is hoping for but it also means building companies
which is quite difficult and not many people want to do people like to work on difficult problems work
on math and try to solve this and then someone else is going to build a company okay thanks a lot guys
what's next for you it's like what's happening with homo how does the roadmap look like when can users
actually try it out as you heard before i'm personally waiting for it to finally utilize my usdc properly
yeah for sure i'll definitely i'll definitely keep that in mind um yeah yeah so i guess like some
the timeline for us is we're currently working on our mvp in tandem with the giza team so we're
working on the smart contracts uh 4626 in tandem with 7540 for some arbitrary arbitrary or asynchronous
logic um aiming to you know we're basically focusing on this and then we're going to aim to like
get our you know core repo under audits we're going under one like major last refactor so we'll see
how long that takes but aiming to start audits end of august and then you know we'll see how long
audits take but i'm gonna guess it's gonna be like two to three months after that
okay
fran how does it look like for you
um on september uh we'll be opening up for everyone these uh verifiable agents to be used and create
automation strategies and what not from a very easy UX perspective so stay tuned for that and
also we will be producing like many super cool applications on that area one of them like being
with Omo but there are like more than 20 in development currently so very excited to
so everyone hear what we have in mind for any applications that sounds amazing the agents to
be used that's for developers or users users everything oriented to consumers okay great
uh i'm good with all my questions i hope it was educational i learned quite a bit i don't think
we have more questions from the audience that could have been submitted with the form so
thanks a lot for your time we won't take it long unnecessarily so let's just get back to
building applications or in your case fryan enabling others to build applications
yes sir thanks for the space thank you guys take care thanks guys thank you for the space
Host
Oasis
from
Oasis Network
