Projects
![]()
No projects found
Investors
New
DAO Treasuries
DATs
Lend / Borrow APRs
Chains
Categories
Spaces
AMA with Dfusion AI: Building the largest Decentralized Knowledge Base with Community Wisdom
Recorded: May 21, 2025 Duration: 1:06:03Space Recording
Short Summary
D-Fusion is set to revolutionize the decentralized AI landscape with its innovative knowledge base and strategic partnerships, while preparing for a token launch that promises to enhance user interaction and growth.
Full Transcription
Thank you. We will probably wait a couple of minutes because we might be expecting more people
to join in the coming minutes.
But in the meantime, if you could do a quick mic check with Patrick,
and that will be awesome.
I'm not sure if it's maybe me,
because I've had issues today with .
But I can't seem, but it sounds staticky to me.
Not sure. Thank you. I'll do a quick mic check from my end.
Am I audible now?
Oh, yeah.
I can hear a lot better now.
Okay.
Yeah.
It's for some reason my laptop's acting funny today.
So luckily I keep both my laptop and my phone, so in case one of them is not reliable, then I have the other one to back up with.
Awesome, so we can probably get started maybe in a minute or two, because I'm seeing more people trickling in.
minute or two because I'm seeing more people trickling in so yeah let's just
give it a couple of minutes I was initially planning to play some music
through my laptop but because it's a mess to kind of get the sound through
we'll just have a bit of radio silence or or maybe we could yeah probably we
could get started with quick intros I can start off by introducing myself
I go by Mr. Potato I'm part of the Nier community squad we run the community platforms
for the Nier protocol as well as we do a bunch of stuff on social media through the Nier Quant
account so one of the things amongst others that we do is we try to highlight
ecosystem projects that are building or starting to build on Nier.
So if there's anything exciting brewing up in the ecosystem,
we try and get in touch with these projects so that those can in turn get
closer to the community.
So there's more awareness about them of what
they're building and some of the exciting developments they are cooking.
So today we have a project which I think many people might not be aware of, but it's super
exciting, which is why we have today Patrick, which is called D-Fusion.
which is why we have today Patrick, which is called D-Fusion.
So yeah, I'll just pass it on to Patrick for quick intros about himself.
Yeah, sure.
Yep, I'm Patrick.
I help build technology at D-Fusion.
I think we're going to get into the product a little more there.
Just some things about myself.
I work on technology there.
I'm also a little bit of a digital art collector.
And I've been working in crypto since 2017, seriously.
I looked at the technology before that.
But around 2017, I started helping build analytics tools for
kind of monitoring, tracing where funds go on chain for research, for lawyers, for law enforcement.
So I was building some of that tooling. And that's kind of what got me into the space. I learned a
lot doing that, because I had just analyzed what everyone was doing
on a bunch of different chains.
Yeah, and I've been building since then
and my latest project that I'm working on is D-Fusion.
So is it safe to say that you were like
an on-chain sleuth at some point?
Yeah, actually, yes.
I had to do a bit of that.
Part of my job, because our team was fairly small.
So that was a team, that company was called Bloxier.
We were acquired by a mining company who wanted to do some analytics, KYC, risk management
type of stuff, right, with their coins.
And so, yeah, we have a, yeah, I had to do some sleuthing.
I had to learn how to, you know, I was just reading all the smart contracts being deployed
on Ethereum, seeing where like ICO funds actually went, looking at things in Tezos, seeing who's
delegating, seeing where funds go around there.
So yeah, I was doing some
on-chain sleuthing for a bit. That was a part of my life in the past. Yeah, I will say it was a lot
harder a few years ago. So now we have tools like Arkham, you've got Look on Chain, you've got
some of these platforms which really streamline, you don't have to dig as much.
It's easier to kind of connect the dots.
But I remember, like, back in the day,
and even Etherscan wasn't as sophisticated as it is today.
As it is now, yeah.
So, yeah, I was always kind of interested, but there was a time, like,
I mean, after a while, I'd always hit the wall, and I'm like,
I don't know, I can't find the trail um but yeah like this it was something i was also interested in um just out
of curiosity but um you kind of i mean because i kind of don't have a lot of programming background
or coding experience like it was hard to kind of decode like some of these transfers and stuff.
And then obviously there's chains where it's a nightmare
to understand what the Explorer says, like Solana.
So there's a problem.
Yeah, so a lot of our job was like just building some of those tools, right?
Just like indexing data, getting into the useful formats that we can query,
um, running whatever ML models or whatever we have on it for whatever limited ML stuff we were doing.
So, yeah. Awesome. Yeah, it's a lot. It's a lot of work. Yeah. And the tools are much better now.
The tools are great. They make life easy. Yeah. It's like, it's now like dummy friendly where
like even a random person could just query up an address and it does kind of like it's now like dummy friendly where like even a random person could just
query up an address and it does kind of it does a pretty good job at like linking which wallets
did it get funded through and like which which other wallets might be associated with it um that
was a lot harder like was at least like i was doing it very manually so uh glad that there's there's a bit of um
there's there's substantial improvement on that and but yeah um i don't want to go off a tangent
anymore um back to diffusion uh i also kind of love your pfp it's uh is it is it an nft or is
it just digital yeah it is an nft it NFT. It's a Mitchell F. Chan.
So also just shout out to one of my favorite artists,
digital artists, Mitchell F. Chan.
He creates, this is Boys of Summer PFT.
He creates, you can play it.
Anyone can play the game if you look it up.
Okay.
I think it's pretty cool.
His art is really cool and I'm into it, so look him up.
Nice.
So, yeah, I think we've got quite a few people in,
so we can probably start dissecting Diffusion, what it is,
why you start building it, and what's all this thing about
building a decentralized knowledge base
sure yeah i guess i can jump in so basically what we do is um we attract and since we attract and
synthesize data into valuable ai assets so that could look something like um on our genesis
platform right now we've been collecting Web3 data.
So we've been building an open kind of Web3 knowledge base that the community can leverage.
So that's like one type of asset we can use.
That's something that you can have your AI query against.
Another example might be something like more specific and business focused like preclinical like kidney therapeutic
deals and advances that are coming out. Maybe you're interested in monitoring that space,
whether you are trading or you're an investor or you are actually trying to develop technology in
that space and you want to understand what's going on. And then another example I would give is maybe you want to collect online arguments to train an ex-bot to troll people, right? And just
be good at arguing, riling people up. Maybe that's a type of data that you are interested in collecting
for whatever AI application you're building. So that's what we do, right? We attract and then kind of curate and synthesize these AI assets that folks can use for RAG, for training in AI. And so we've been
fairly focused on language data starting out, but we have other types of data that we're
branching out to. I think the long-term goal is to kind of, is to build the incentive system that fuels kind of the AI-driven
global data economy that is developing. Yeah. Yeah. I'm kind of curious, like, is there any
particular reason why you decided to build this? Like, is there like a moral reasoning? Yeah. Yeah. Yeah. Yeah. So two things. One is before this, we were actually
building, so leveraging some of that background I had in, so my co-founder and I, we'd actually
been working in web two for a little bit. We were building out insurance products and helping people
sell their like cyber insurance, things like that.
And we were actually getting some referral business where folks were asking about having
crypto assets under management, what kind of tools they can use to track and trace funds
or to produce evidence that they're custodying things correctly.
And so as we saw some of that business, we decided to build
a knowledge base combined with like around this time, I think 3.5 had dropped, GPT 3.5 had dropped.
And we decided to like have kind of like a searchable, interactable database. You know,
you can interact with a database that'll like help you hook into tools for tracing and tracking crypto and kind of understanding what are good principles to follow, what resources are out there for custody and wallets, doing things like that.
So we were curating this kind of knowledge base ourselves.
And what we realized was two things.
And what we realized was two things.
One is that we didn't, from a values perspective, we didn't want to be the only arbiter of truth
because we were picking, you know, just based on our business development, that would determine,
hey, who are we recommending you go look at for, say, tracing stolen funds or something
like that, right?
And so that would just be based on our business relationship.
So that would just be based on our business relationship.
We're the arbiters of truth or we're telling you who to look at for that.
And we thought, hey, in a better world, everyone knows all the tools that are at their fingertips and at their disposal.
And then we also thought, hey, these models are getting pretty good.
good, there is a long tail of problems that previously folks wouldn't invest maybe in solving
that now we can build kind of guides and agents around. And so we thought, hey, that will require
these kind of niche data assets that will require an economy where people can quickly describe
the type of data
or knowledge they're looking for, and then be able to attract and get that data to build
their applications. And so we decided, let's build the tooling to do that, to both take
the power out of our hands as arbiters of truth, and to enable people to build their own systems as the market evolves, right?
Yeah, that's great to hear.
Like, I was just kind of curious on, like, why the shift to kind of build diffusion.
So the other question I kind of had is around decentralization, right? So most of the AI
that's out there, where it's models, or even like solutions, they're not decentralized. So the likes
of open AI, et cetera, et cetera. So it might be like for folks in the new ecosystem, like we've been advocating a lot on why it should be decentralized.
But maybe in terms of like data, if you could shed some light on like why is decentralization so important in the way we kind of store and verify our data.
so important in the way we kind of store and verify our data?
Yeah, I think that there's two.
Yeah, I think there's two things here.
So or let me back up a little bit.
So I do think we will continue to have kind of centralized
like groups and labs and companies building
like frontier AI technology, though there are folks challenging that.
There are folks in the BitTensor ecosystem
who, for instance, are trying to build and train frontier models
using decentralized approaches, right?
So we do see some of that.
I think a lot of it is kind of like this,
you'll see the values-driven side,
which I think we've talked about a little bit, which is kind of like, we want kind of a pluralism around this, right? We don't want the arbiter of truth to be like, just Sam Altman decides what's true or whatever.
We do care about, you know, what data is going into our models.
If we're using some system where there's RAG or like a perplexity behind, we're using perplexity or something.
What sources are we trusting?
Who are we believing?
You know, I think a common thing you'll see on here on X is somebody will say, at Brock, is this true?
And you will see varying results. Sometimes Grok is very accurate there.
And then sometimes you'll see it just kind of, it's guessing. And so it makes something up and
that could actually be harmful. So I think people are using these tools and they will come to rely
on them more. More companies will use them, more builders will use them. And so from like that values perspective,
I think having like a pluralism there is super valuable. And then from, I also think there's
a practical and opportunity side. Like I said, I think there's like a long tail of problems.
So let me just give an example here. You know, people have probably heard of Vibe Coding in here. I know someone who, she's a writer. She writes erotica, actually. But she actually has an AI pipeline that she uses to edit her writing, to help her edit, right? for going back and forth and helping her do edits and analyze plot points and things like that. And even do some market research.
And she was copying and pasting stuff.
She actually like vibe coded her own platform together where she's automated
those pipelines. So, you know, she's not somebody who coach, she's,
this is a writer.
So we now have tools where folks can solve problems that they didn't use to solve.
I think that when it comes to data, there's a huge opportunity to build niche data sets that power.
You know, if we have verifiable, trusted data that can power these different niche use cases that we previously wouldn't have like the
ability to invest in and the AI has made possible now. I think that will be a super valuable thing.
So that's the other part from decentralization is I think that we can enable like that more kind of
long tail data market for folks building these new tools.
Yeah.
Yeah.
I definitely want to touch upon what you said with people,
like it's a growing trend of people like tagging at Grok or even
for Plexity is like a popular option where you're kind of fact
checking or like getting more insights into something which someone else has tweeted about.
And the question I kind of, and you kind of alluded to this, right,
that there's often times where it hallucinates or it kind of get more accurate and quality responses when it comes to getting AI involved?
Yeah, so there's a couple things.
One, yeah, I guess there's multiple parts.
So one, I'll just say we do also operate.
So we're trying to interact with everyone.
We're building apps on top of,
we think there's a ton of deep in out there.
There's a ton of interesting decentralized technology
out there.
We should be building applications and protocols
on top of them and leveraging them
and kind of synthesizing across these networks.
So I mentioned BitTensor earlier,
we do have a subnet in that ecosystem.
And that is its focus, actually, on corroborating
or refuting statements, basically agreeing, disagreeing with things, and then providing
evidence on both sides of that. So we're interested in doing that. But the answer is that it's
actually hard to do this. And it depends on the domain, and it depends on the type of data you're collecting.
So in the Diffusion model, like what we're doing in Diffusion, we have this idea of what are called subnets, where you can describe the type of data that you're looking for.
Like I said, it could be arguments, and so you're looking for. Like I said, it could be arguments.
And so that's, you're looking for chats online
or it could be data about preclinical deals, right?
And so those different types of data,
you'll want different types of checks on provenance,
on the format of the data, on the content of the data.
And so our subnet system is a system on the format of the data, on the content of the data.
And so our subject system is a system where we'll actually use AI to validate that the particular rubric for the data you're looking for is met.
So we're actually leveraging AI there.
There's an AI task, verifiable AI task.
And then we also have kind of like this semantic linking that we're thinking of.
You know, think like querying perplexity or whatever. Do we have evidence to back those things up? Do and then surface those so that users consuming that data can understand, ah, yes, you know, somebody checked that this is
actually from this research journal, right? Here's cryptographic proof that the data was pulled maybe
over a TLS connection or whatever from like PubMed or whatever, right? So whatever those data needs are,
that's kind of what we're making sure we can support
all of those different kind of needs for different data sets.
That's great.
So I guess like one of the core motivations
to kind of build this on Web3 is to kind of figure out how to
how to solve the coordination problem right like you want people to kind of
act as contributors and like people essentially if I'm and correct me if I'm
wrong it seems like the goal is to kind of build some kind of marketplace where you have like both sides interacting and making sure we have accurate and reliable data when it comes to AI, right?
Yeah, I definitely think, yes.
So I don't know if I would describe it exactly as a marketplace, but yes, right?
Like market driven, kind of like both sides of the market, right?
Like you can think of it as like the supply side of like there's folks either sitting on data or they could be contributing data in aggregate.
You could get like some like wisdom of the crowd type of stuff, right?
stuff, right? Community wisdom type of information. It could be that you're a particular entity and
Community wisdom type of information.
you're sitting on very valuable data set. And then there's obviously the demand side. And we
focus heavily on that demand side because we want to make sure we're solving business problems. I
think right now, you talked about verifiability too. There's a lot of, as I mentioned, D-Pen platforms, DAI platforms out there.
And if we can, that means it's a great time if you want to build an app, right?
If you can solve a business need, there's a lot of fairly mature tooling out there to build on top of.
of, yeah, so, yeah.
Yeah, so, yeah.
Yeah, so, yeah, like how would,
so let's like look at the supply side.
You said there's maybe like data suppliers,
like entities who are sitting on like valuable data sets,
but also there's this
contributor who maybe like are creating that data
like making sure it's of good use.
So what roles would they play?
And the other part would be is how can people participate
as contributors within Diffusion?
Yeah, one simple way to participate is you can literally just go and upload data.
So we have a genesis.diffusion.ai.
We have a genesis page where you can contribute.
Right now we're giving out, we have like a point system.
We're giving out points and that gets like folks access to like whitelists. And I can't, I think I'm not technically supposed to make any promises or anything, but like rewards down the line, right? We want to reward our community members, things like that.
we have that you can go in there and you can actually see some of these subnets and some of
these rubrics and you can upload data and you will uh you will receive at the moment it's a
point system you'll receive points for um those contributions and we've had a ton of people
contribute through there we've over like 650 000 users contribute data on there and we've got um
just tons and tons of data that's's mostly on building that community knowledge pool around kind of Web3 and AI.
But we have other topics in there
and other subnets in there, as I mentioned before, too.
And yeah, so that's an easy way.
You can just go and you can upload data
in the form of files or links.
And that's an easy way to get involved.
Yeah. Yeah. So all of the, like,
would you say like most of the people who are,
who've joined the platform are like contributors or like what,
what's kind of like the split?
Yeah, it's been, I would say it's been, if you just raw numbers,
majority are contributors just raw numbers, majority are contributors.
Just because, yeah, we've got a lot of individual data contributors.
Some of the, you know, I think it still follows like a power law.
Like we have a couple, we have like some percentage of users that contribute most of the data. But we have a lot of, yeah, we have a lot of contributors.
contributors. And then on the other bit is we're working. So our org has been doing work, kind of
And then on the other bit is we are working.
doing business development with folks who want to test out these systems or trial out these subnets.
So that's also going on the other side that maybe like you see a little less publicly.
publicly maybe you just see like the subnet pops up in that contribution portal for you
Maybe you just see like the subnet pops up in that contribution portal for you.
um yeah yeah it's always like you have a small subset of users who are kind of like
your power users or your power members they kind of uh contribute like the vast majority of volume
or or whatever you're doing so um it's I've seen that like across like multiple platforms,
whether that's like Web 3 or Web 2.
So you mentioned points.
So I think on the website,
I read something about like knowledge injection points.
Is that the points program like you briefly touched upon?
Yes, that is.
Yeah, yeah.
So that's, again, those will be rewarded based on like,
hey, are these meeting the criteria of any of these subnets?
Like the validation process by that?
Was that information corroborated?
How unique is that data within that subnet?
And the different subnets have different uniqueness properties that they are interested in.
And so, yeah, that, that will be points.
One thing we've let some users do with those is if you have enough points,
we open up like some of the search tooling on like, you know,
beta access to the search tooling.
We also are letting some of those people vote on and propose subnets that
they'd like to operate or collect data on and get access to. So we're doing that at the moment.
And then, like I said, down the line, the idea is we're representing that as, you know, we're seeing
which community members are really participating and we want to be able to reward those people.
Awesome. I might want to ask like a cheeky question. Usually when you have like points
and points program, it eventually ties down to a token at some point. Is it correct to assume that that will be the case?
I don't know what I can say there. What I can say is that we intend to, people who are helping us in our system, we intend to reward that and respect that people have contributed to our ecosystem.
But yeah, I'm not sure what I can or can't say there.
So sorry if that's not a full name.
Yeah, don't want to put you in a tough spot.
But it does seem like the points program will carry a significant amount of reputation within the platform.
So that's really good to know.
Now kind of like finding the intersection between Diffusion and Nier.
So as you know, most of the people who join our spaces are community members.
And they're always looking out for opportunities both to support ecosystem projects, but also to kind of get involved in them.
So as a Nier community member, what can we do with
diffusion? Yeah, let me just say some things about near too. So I mentioned that we're kind of
working with multiple protocols and like deep end providers and things like that. So like one of the
things we really like about near is kind of this focus that's existed on like MPC, like this multi-party
compute technology, cross-chain technology seems fairly mature and pretty cool on Nier. So
actually synthesizing an application on top of all of this new stuff that's out there that you
can utilize,
that seems like it's a great place for application builders.
And from what I've seen,
Nier is also doing a good job of like running accelerators
and talking to cool AI projects who want to build.
So yeah, just a cool place to be building, I think.
And so yeah, folks want to get involved.
I think, you know, you can jump in our Discord.
That should be linked on our you'll see like the Diffusion AI account is in here.
You should be able to find it there.
We also have you can just contribute if you'd like to just Just go to the Genesis site and upload files.
The other thing I would say is if you are somebody building in the space or you're working on Intense or you think that your agent would like to have access to certain kind of data, reach out to us.
Reach out to us in Discord. Reach out to us on Twitter, and we will, we're happy to build there, right?
That's what we're trying to do. We want to collaborate with everyone on finding folks
who have demand for this data or who would like high quality data for their applications. So
those are ways you can get involved. We do have, if you want to check out our GitHub, we have
open source repos for across multiple different protocols, as I said.
So you can see some of our code and tooling there
if you're somebody who wants to do that.
Yeah.
Yeah, one thing that often gets requested by builders
is access to data from Twitter or X
because it's usually too expensive
for new devs to kind of pay the Twitter API.
So they often rely on like workaround solutions.
And there's a few projects I've heard about
which kind of facilitate access to kind of data on platforms like Twitter.
Is that something you guys also try to provide in some way?
So in terms of those public sets, we haven't run as much of that.
Someone could.
We could set up a... We could
help. If someone's interested in setting up a subnet around that, happy to talk and we
can get that going. We do see that there's a lot of people focusing on that data, as
you mentioned. One data set. So we also work with the VANA ecosystem. We have what's called the data liquidity pool there.
There we've been gathering Telegram, like private chat data that you can do like kind
of anonymized inference against, run tasks against.
We're currently working with VANA on building out that system.
So if you have ideas around, if you like TEEs
and building that kind of stuff,
talk to us about that as well.
But that's something where maybe if folks want
either insights from, hey, what are people saying
when they're talking with just like a couple
of their friends, or you want to build on that type of data,
that's actually a dataset we've been gathering.
I think we have like millions of chats in there now
from people's contributions through that system.
So private Telegram data, we plan on expanding that out to WhatsApp
and some other messaging platforms as well.
And again, the idea behind that data set is mostly,
instead of what are people saying in public,
can we safely pull insights out of what do people say in private?
Yeah, that's something I haven't heard a lot about.
Usually it's mostly like getting data through public platforms, right?
Where like data is fairly accessible.
But I think like Telegram is a very untapped resource.
Like private chats, like there's a lot of potential.
I don't think I've heard of like any project
like seriously looking into it,
but I do see it as something that has
potentially a lot of scope.
So that definitely sounds interesting.
And WhatsApp is interesting as well.
But one thing I kind of think is annoying on WhatsApp
is that they've kind of forced people
to have meta AI inside the app,
which is a huge turnoff.
But I've been trying for like days to figure out how can I disable it, but I can't seem to do that.
Yeah. Yeah. Maybe paying attention to where privacy is still relatively preserved is a good
way to make sure you're getting good signal from those kinds of data yeah and so and yeah that that idea of um targeting that data you know i think there was
some story sorry if i'm going on a tangent here but there's a story of a guy who um uh and i don't
remember all the particulars but there was a guy who commissioned i think he spent like a couple
million bucks commissioning uh his own polling in the united states for the U.S. election. And the way he
asked, apparently, supposedly the polling was done where he asked folks, who do you think your
neighbor is voting for? And it turns out that folks are more willing to give accurate data of
who they're going to vote for, but also what's going on in their community
whenever they're asked the question that way, instead of being asked who they're voting for.
And, you know, hard to say, like how much that was actually signal versus noise, but apparently
his polling was fairly accurate. And so he did a great job on betting markets and polling market.
And so he did a great job on betting markets and poly market. So I do think, you know, seeing what people say in their like, in like smaller groups and friend groups and in smaller channels, probably is a, you know, you can have some valuable insights there that you wouldn't get maybe so much on public channels.
sites there that you wouldn't get maybe so much on public channels.
And, you know, if we do that, if we're gathering that kind of private information, you know,
it has to be secure.
You have to make sure that, like, you know, you're not just running inference or you're
not aggregating data on someone's machine who's just logging all of that user data,
right?
You have to have good kind of trust properties.
You have to have good privacy preserving properties there
because, yeah, that's more sensitive information, right?
100%, which is why it's been more of a challenge
to figure out a way which, as you said,
is privacy preserving.
And it's kind of like a safe sandbox, so to speak.
But, yeah, I think it is very interesting
because you would speak very differently in a private Telegram
versus what you'd be tweeting or like creating posts on Instagram,
right?
Like how you approach things like this.
There's a lot of nuance, like maybe like there's a completely different conversation
style.
There might be more joking around.
There's a lot of like, it's the way to kind of analyze private data now that I'm thinking about it is probably very different to like the conversations or the social media posts we see like on platforms like Instagram and Facebook and Twitter maybe.
Yeah.
And I think what you pointed out of like you might speak differently. Like forget, you know, I was talking about insights, maybe aggregated across a bunch of people. But, you know, if you're training AI models that are supposed to be conversational or feel more natural, you know, if you talk to, you know, if you if you talk to any of the big AI models now, and, and you tell it to try and be cool, it comes across as not very cool, right?
It's very, it's kind of trying.
Yeah, and maybe trying too hard.
Yeah, maybe trying too hard.
And so, yeah, even from the perspective of like training conversational AI, having that
data at your fingertips, maybe that's useful in like fine tuning or training, training more conversational ais training ai that can be a
little bit cooler yeah um so coming back to kind of um centralized and decentralized ai so right now
um it does feel like over the last year or so, the decentralized AI ecosystem has evolved
a lot and surprisingly a lot.
We've seen the likes of DeepSeek, even NIR's double down on AI.
We've got NIR AI trying to build one of the largest language models, kind of decentralized, but
kind of wanted to take your takes on where do you see this space evolving in the next
two or three years? We've seen like progress that maybe very few people would have thought of like a year back.
So yeah, since you're more involved in the day-to-day of the AI world
and specifically the decentralized AI world,
yeah, if you have any insights
on like how fast could this thing continue to grow?
Yeah, I mean, I wish I knew
exactly how fast things will change but um
but you probably have a better idea than i would yeah yeah i i would say i would say um i think
probably uh i might have said this earlier but i think probably we'll still see like the large
orgs um do pretty well right at building like these models, state-of-the-art models,
just because of the resources they're able to pour in.
But another thing we're seeing happen as well is, you mentioned some of those competitors
are going to open source things.
We have reasoning models kind of in the wild now, thanks to like DeepSeek's paper
and they're open sourcing their weights
and all of that, right?
And I think the other thing we're seeing
is that hardware will get better for inference
and we'll be able to run smaller models
that can do, there's already,
there's a ton of people, I mean, go to Hugging Face,
there's tons of people building,
hey, can I build a smaller model that does more or like a small model that performs as well as maybe one of these more general LLMs? Maybe it performs I think, you know, if we can do things like take reasoning
and make it work and run on just about anywhere, and it's fairly ubiquitous,
that world, the type of stuff you can build in that world looks pretty different. And
yeah, so I think there's like a lot of opportunity there. I know I mentioned like long tail applications, things like that.
You know, folks, you know, I think of things like, you know, WordPress existing and then folks actually being able to put websites together.
businesses that wouldn't put a website or like a seamless or a grub hub where now now restaurants
might have some analytics and some insights into like what kind of pricing they want to do that's
a system that you know probably your local your local restaurant wasn't going to hire a data
science team so I see this with AI right like where it's going to become more ubiquitous, there's going to be like a huge economy around it. And so what I see happening is kind of, there's opportunity in like the Web3 space, decentralized space, to kind of build the economic rails for data exchange and stuff to happen
for that kind of like those long tail use cases
and for like the wider global like data exchange.
So that's why I'm interested in data there.
I think it like matches from like a business
and technological perspective
and from a values perspective,
as we talked about in terms of not having like one person
be the arbiter of truth.
So I'm hoping, I think we'll see a lot more of that.
I think we'll see a lot more applications
that not just use RAG or AI behind the scenes
or use a chat bot, but we'll see a lot more applications
where they're actually internally deploying models
and using them for other tasks that aren't just chatbots.
Yeah, I do agree with something that you've mentioned,
and I'm seeing a lot of,
is ever since we kind of had DeepSeq release their model,
like there's been a shift to kind of optimize stuff, right?
Like to do more with less. And I think rightly, as you said, like there is indeed like a lot of
these experiments or at least like attempts to kind of optimize smaller models so that they're
better at like specific tasks. And like, I kind of think maybe some of these models will then turn,
be spun up as agents.
So you have an agent which is running this specific small model
for a specific use case.
And then you have this agentic world,
this agentic world, which we've been talking a lot about in the near ecosystem.
which we've been talking a lot about in the Nier ecosystem.
So I definitely see more of this, at least like in the next year or so, given that like there is,
there is now some kind of belief that you are able to do more with less. But yeah, on Diffusion, maybe we can probably talk about the roadmap. What have you guys done so far? What are your plans to kind of scale the project in the months to come? And then if there's any milestones that the community should kind of keep an eye out on?
any milestones that the community should kind of keep an eye out on?
Yeah, those are good ones.
So a couple of things.
So in terms of what we've built so far, I think I mentioned,
check out our Genesis system.
We have just tons and tons and tons of data in there.
And people are always contributing.
So please keep doing that.
We appreciate it. And then, as I mentioned, we're interacting with multiple systems.
We have millions of check contributions for our DLP on the VANA ecosystem.
We're constantly trying to improve the incentive mechanism for our evidence corroboration subnet.
So we have all of those things going on that folks can check out.
And then coming up soon, so we do have things for each of those coming up.
But I think a big one coming up soon is we'll be doing a limited release of subnet slots, like doing a small subnet slot sale there.
And so keep a look up for that.
That should be coming like fairly soon.
And then the idea is to take those
and let people interact in a testnet environment
where they're actually able to like own their own subnets or delegate them
out if they want to rent the slot or whatever. So that's something to be looking at coming up.
Hopefully we've got some new, we've got some cool announcements there on both the tech side of how
we're doing those validation processes in a testnet and on the subnet slot side.
So I'd say that's the next thing to keep a lookout for.
And pay attention to the subnets that we release.
We're trying to launch like a couple of subnets every week.
And those are generally around businesses and power users who we've been speaking to who want to gather data sets that would power their applications and that they'd be willing to pay for.
So keep an eye on those as well.
as well.
Yeah.
Yeah.
Awesome.
Awesome.
And yeah, maybe like as my final question, which I generally tend to ask a lot of projects
is, and I think you briefly mentioned about this earlier, is how can people get involved,
whether that's contributing or learning more? Like, is there some kind of
like a checklist of like things you should kind of look into? Like,
what should people get started with first, et cetera, et cetera?
Yeah, yeah. Really easy ones is to like, you know, join, you know, join the discord,
is to join the Discord, check out our X account,
and then contribute the Genesis Diffusion AI site.
Just go in there, check it out, contribute some files.
On the DLP side, we link to that from our site as well.
You can go and contribute your chats if that's
something you'd like to do if you'd like to read up on the docs we have some faq and link to the
docs on how that data is secured and then from the and then and then getting more yeah getting
involved down the line I think keep an eye out for any like uh near news
if you are someone who is trying to build uh some integration in intents or you're trying to build
uh agents that can be utilized for that stuff for helping power intents uh and you want um
a particular data set you want to trial it with us we're always happy to run experiments with
people so reach out to us uh there and the easiest way would be probably in our Discord there,
or just try and ping us on X, and we'll try and keep an eye out for that as well.
Awesome. Yeah, I think the combination of intents with AI is going to be a very interesting one. We've already seen some of these agents,
which Nier AI has been working on,
which enable Fiat on ramps and stuff.
But I think that is just scratching the surface,
to be honest.
There's a lot that can be done
if you incorporate data-based AI agents that can maybe like...
I'm not sure what the possibilities are, but I think it could be really interesting.
But yeah, I think we've got about 10 minutes or thereabouts for community questions, which
we usually take towards the end so
we've got a couple of people who requested to speak so I'll bring them up
on stage and we can we can then hear from them and what what they want to ask
you guys so I brought up two people on stage.
They should be... Okay, both of them are on.
Maybe we can start with Leotone first
and then Sushant.
Yes, hello. Do you hear me?
Yeah.
Yes, thanks for MA.
It's a totally new project actually for me.
And I have added, have followed your socials to x, discord and so on, see that you have
some additional activities for this social.
Thanks for it.
And my question related to the MA is like, what is your vision for interoperability between diffusion
and other AI, the pin protocols?
And could diffusion become like analog
trust Oracle layer across Web3?
Thanks.
JOHN MUELLER Yeah, that's a good question.
I think the answer is yes.
So as I mentioned, we already work with,
so like on GPU side, we actually supply some for like our validators and stuff,
supply some of that GPU from deep end
so that some of that comes from like the Aether folks.
So Aether powers some of our stuff there.
As I mentioned, we have, we work with, in the BitTensor ecosystem, we work in the BANA ecosystem.
So we have, we are talking to all of these chains.
That's why we like Near.
We like being able to talk to everyone.
We like kind of this intense system where you can kind of synthesize across these different protocols.
So that's a part for us. And yes, so in terms of can it become like maybe an oracle for these
things, I think that is the idea is that as we're building and solving these business needs and
saying, okay, here's the best solution for this piece of validation. Here's the best solution for this other piece of verification, whatever, right?
We are designing the incentive system,
and this is what we're hoping to test out in our testnet and stuff,
is saying, hey, here's how we can take evidence from all these chains,
and we can kind of
validate that we have from these different ecosystems, run these tasks or gather this
data from those places. And so we can kind of give you like that chain of proof that we've done,
we've done the work to provide the data that people demand. So yes, absolutely that is the idea is
we're synthesizing across
these different systems and
we would act in some
kind of like oracular
way there. Yeah.
Okay, thanks. I understand. Thanks
once. And as I told
I have followed you, so hope
to see it with my own eyes thanks
yeah awesome awesome thank you cool uh maybe yeah i think up next uh we have sushant um if you want
to unmute yourself and um speak you can now do so hello guys thank you for such an informative image.
My question is how does the OpenFusion kernel ensure the accuracy and reliability of the
knowledge contributed by the community?
Got it, yeah.
So that's in reference.
So there's multiple parts.
This is a multi, the way it does that, it does that in multiple steps.
One of those is through...
Right now, we're running a federated system.
We are moving that to an on-chain verified system.
But one of the pieces of that is that we actually take the subnet's criteria,
take the subnet's criteria, and we actually utilize AI to check if those criteria are met.
And then we have other validators reproduce that work and then come to agreement or disagreement
on whether those criteria are met. So that's one layer is like using AI for validation,
which is something we, you know,
previously would be fairly hard to do,
but at least with language data,
you can do that.
You can actually do that fairly well now
with reasoning models and LLMs.
And then there's the other layer of veracity of claims,
what kind of evidence supports it and backs it up.
And that is where we use that evidence gathering system
that's like on the bit tensor ecosystem we're using there
to kind of validate those claims.
And then even further down, for specific data sets, we have other types of things too.
So for instance, verifying the provenance of data using something like a ZKTLS or whatever.
So there's actually multiple stages in that process and different datasets will use different bits of those stages.
And we try to make it easy so that somebody who's defining one of these subnets
doesn't have to know all of these details behind the scenes,
but can describe the type of data they're looking for.
And then we're able to kind of help them appropriately
determine which of those systems should be applied
in verifying their data.
Thank you for such a detailed explanation.
Hope you have a good question.
Yeah.
Awesome.
Maybe we can go to Silver Surfer next.
If you want to unmute yourself, go ahead.
Hey there.
I was waiting and going to the website.
And I think about the VFSN tokens incentives
to the contributors.
So I was just wondering how you guys differentiate
between the high quality intellectual data
rather than incentivizing the people
who contribute low quality data to your platform
with a flux of data.
You know, when there's like point system or token system,
people find a way to exploit it
through providing a flux of data which is not
that quite uh informative or kind of you know usable so how do you differentiate or you synthesize
or you try and game the system right yeah yeah yeah yeah so to stop any kind of exploitation
and keep it uh secure and you know, reliable. Yeah, yeah, yeah.
So I think this is kind of related to the last question.
So I think you mentioned at the beginning, like VFSN, VFSN specifically,
the way the DLPs with those DLPs.
So that is the token that the contributors on the telegram,
on that telegram data get.
And again, the answer comes down to,
it depends on what the criteria for the data set are.
Some of that is that we, as I mentioned,
we'll use AI to analyze the data,
if it's supposed to have certain properties,
if it's supposed to be in certain format, things like that.
And obviously people try and game those models as well, right?
So like the solution, there's no actual solution to solve this in 100% of cases.
But what you can do is for a particular niche,
determine which criteria must be met in order to build a good data set.
So an example of that would be verifying, for instance,
that perhaps data comes from a particular domain.
Like if you have like a research article and you can prove that it comes from, that you pulled it from PubMed using like TLS data, that's something that we do.
thing that we do. Same with like chat data, right? Private chat data checking that that's like a
Telegram user, that the data actually comes via the MT Proto protocol from Telegram. So there are
like hard checks. Some of them are like more hard checks. And then some of them are these use
reasoning models for verifying like the criteria and then checking against the existing data for
things like similarity or uniqueness properties. So it's actually just that there's multiple pieces.
You can't, there's not a silver bullet to attack this problem. And even as we do this, we discover
different clever ways that people try to find to game these systems. So it involves like constantly
updating like that incentive mechanism, constantly updating like our understanding of which systems
need to be in place to get reasonable data for these niches, which is a very, it is a hard problem
to do, but we're trying to do it
yeah thank you for your answer and like i was reading in the facts section that
there would be a documentation about the data score of individuals so
like can i uh is it available right now i can read somewhere or is it in the
process sorry which uh which is this this is
documentation on uh how how scoring is done.
Yeah.
Yeah.
So like there was a question on the side that how my data like was scored and
there was a mention that there would be documentation about it.
So is it already live?
We can reach somewhere.
Is this on a, is this with regards to VFSN or is this regards with the Genesis data contribution?
Because those will be a little bit different.
Yeah, it's about the social to DLP.
The Telegram data.
Yeah, yeah.
I think we have docs on that page.
If we don't, let me take a look.
We also link to the GitHub.
The GitHub should have like a readme about that. And you can actually see the task that runs the
proof there, in there. Now, obviously, reading code is not what everyone wants to do. And in
general, the way that one works is you will be scored on
a curve based on like number of participants. We don't want a large number of participants
in a conversation. And like the number of chats and the amount of fresh data, because we're able
to keep track in like a pretty hard way of which conversations have been contributed and not so far. So new conversations, obviously, like validity of it
coming from an actual Telegram chat and user,
and then actual size of folks in the conversation.
But right now, we're casting a wide net
for those conversations.
We just want those small channel Telegram combos.
So you can take a look in there but that's that's kind of what we
use at the moment thank you for your answer and awesome listening to you looking forward to what
you've been for thanks um yeah we uh thanks for the question uh we're actually up on time, but I don't mind going another five, 10 minutes more
if that's okay with you, Patrick.
Yeah, I do have to hop to some other things,
but I can do like another question or so.
Yeah, let's take one more.
So maybe, Charon, if you want to unmute yourself,
you can go ahead.
Okay, if you want to unmute yourself, you can go ahead.
Hey, thank you. Good evening, everyone. So my question is, how can we as the community participant in running the nodes for network validation?
Yeah, that's a really good one. So that we should be opening that up soon.
So some of that I mentioned, like we have our like subnet slot sale coming up so that there will be opportunities there.
But stay tuned.
We have some.
So, OK, actually, a couple of things.
One is if you want to get involved in the evidence gathering, you can already if you've got some technical expertise on it, you can already check out on our BitSensor subnet.
You can actually register your own miner against our subnet.
Our subnet is subnet 70.
You can register a miner and run code there.
So that's one way to do it.
We have some open source code on our GitHub for running that.
So you can mine on that network and you can compete with the other miners on doing a better task. You can game us and then we learn to be better. You can earn some tokens
while teaching us how to write our incentive mechanism better on that platform. And then on
the... And then in terms of core protocol diffusion, running these validator nodes, I would say keep
protocol diffusion running these validator nodes, I would say keep an eye out. We have some tech
cooking. There should be some announcements coming in the not so distant future on how you can get
involved in running that in our testnet. Yeah, so that's coming soon, hopefully.
Yeah, thank you. Thank you for answering my question. Yep. Awesome. Yeah, so thanks everyone for joining. Unfortunately, we're up on time,
so I'm not sure if we can take more questions, but since this has been an interesting space,
maybe later down the line we could have a second space where hopefully we can talk about more of the
have a second space where hopefully we can talk about more of the recent developments
on Dfuse.
So it's maybe like it might be like we could follow up with another one if there's enough
interest.
But yeah, thanks, Patrick, for taking the time out to speak to your community.
And thanks, everyone, for joining thank you everybody
yeah all right um yeah um bye for now bye everybody Það er það er hann.
