Latest / Agentic Conversations (formally mlops.community) / Collective Memory for AI on Decentralized Knowledge Graph // Tomaž Levak // #285

Transcript

Collective Memory for AI on Decentralized Knowledge Graph // Tomaž Levak // #285 · Agentic Conversations (formally mlops.community)

0:00
Hi, I'm Thomas. I'm the founder at Origin Trail.
0:05
My coffee depends on the time of day.
0:07
I'm really Italian, so if it's morning it's going to be a
0:09
cappuccino. Never after 11.
0:12
So between 11:00 and 12:00 to 1 you might get me with a
0:15
macchiato, but after one is going to be all espressos.
0:20
Knowledge graphs, knowledge graphs.
0:22
Knowledge graphs is what we're talking today, but this time
0:27
decentralised knowledge graphs. If I may, I'm your host
0:32
Demetrios, and welcome back to another edition of the Envelopes
0:36
Community podcast. We're rocking and rolling all
0:41
about a gigantic knowledge graph.
0:43
That is the best way to keep things transparent.
0:47
Talking with Tomas and let's just get right into this
0:53
episode. If you want to get started and
0:55
set your own node up, go and hit up Tomas.
0:58
They've got some cool things brewing at the Origin Trail.
1:04
And for those that are with me on Spotify, let me serenade your
1:11
ears with a little Spotify hit that I have been looking at.
1:17
Let me share with you some of my favourite sounds.
1:24
A new single from one of my favourite bands Parcels.
1:36
They just came out with an EP, 7 different tracks.
2:27
man, knowledge graphs, they are so important these days.
2:31
So tell me about what you're working on.
2:35
And Congrats on V8. I know that is huge.
2:40
You, you guys just released that and what exactly are you doing?
2:45
Yeah, I mean, thanks. That's a, it's a loaded
2:48
question. I'll, I'll give it a go.
2:51
And knowledge graphs definitely are a huge part of that match.
2:54
So yeah, I'm, I'm one of the founders of Origin Trail and
2:59
what we've been kind of building out for the last decade or so
3:03
where different components that ended up being a decentralised
3:06
knowledge graph. So here we are.
3:09
We're talking about the good old fashioned AI, the symbolic AI
3:14
where we feel that if you do it in a decentralised way, you can
3:19
unlock so much more power, especially when it comes to
3:23
stuff like data ownership, when it comes to stuff like
3:26
verifiability. And for us, it was really kind
3:29
of helping us to, to build out that vision or, or the, the,
3:34
the, the core believe that if we have somewhat more transparency,
3:39
it, it would be an inherently a good thing.
3:42
And that is much more easily achievable if you don't have a
3:45
centralised entity in the middle trying to grab a hold of
3:48
everyone's data, right. So if you're decentralised,
3:51
knowledge graph is a perfect middleware, like a common
3:54
playing ground where folks can connect their systems, their
3:57
data into into this shared knowledge graph, global
4:00
knowledge graph and build cool stuff on top of it.
4:06
I think you told me there's like neighbourhoods within the
4:09
knowledge graph, right? Or 'cause it's you can have a
4:12
global knowledge graph, but then you've seen that there's certain
4:18
areas of the name of the knowledge graph that are getting
4:21
more populated. Yeah, for sure.
4:23
So you can, you can even imagine it like as a knowledge graph of
4:27
knowledge graphs, you know, a decentralised knowledge graph
4:30
where you would have these neighbourhoods which are
4:33
populating with knowledge graphs around certain topic.
4:37
Let's say we actually call these paranets, so parallel networks.
4:41
And you can imagine that this global knowledge we have, the
4:44
decentralised knowledge we have being just an endless kind of
4:47
assembly of these paranets coming together.
4:51
And these paranets, they'll be your neighbourhoods which have,
4:54
let's say, common set of rules that we define.
4:58
Let's say you and you and I want to start a paranet on on the ML
5:01
OPS podcast and we'll say, you know what, like we want to see
5:05
contributions with these ontologies, that type of data
5:08
structure, because for us, that's important when we'll be
5:11
running our system like a solution on top, which is like
5:15
an AI powered, you know, system where it's going to be some some
5:20
some powerful ML that we want to perform on top.
5:23
We're going to know what to expect.
5:24
We're going to know what ontology we're getting their
5:26
contributions in and then for anyone can kind of contribute
5:29
that we allowed to or did we want to even motivate still with
5:33
with incentives. So this, yeah, neighbourhoods
5:36
are are something that's important, important concept,
5:40
especially because knowledge graphs can be much more powerful
5:44
if we kind of fuse the things that they're natively built
5:48
together with that lycontologist present.
5:51
Yeah. So you how, how many layers deep
5:55
have you seen it going? Yes, it feels like it could go
5:59
very, very deep, especially if you're defining something that's
6:04
quite complex. Yeah, no, I mean, I'll say that
6:08
for example, you, you have the DKG today's being used by some
6:11
of the, let's say, enterprise users.
6:14
So you'll have, you know, Fortune 5 hundreds or Swiss
6:17
Federal Railways. And a lot of their use case,
6:19
they're actually using the DKG but with private data.
6:22
So their parents will be assemblies of this knowledge
6:26
assets, which are atomic units. So they're like entities in
6:29
graphs, right? And these knowledge assets are a
6:32
part of these paranoids, but they're not visible to anyone
6:35
because they're hosted by their notes.
6:37
And because they only, they say, OK, we want to, we'll publish
6:40
the the things to make them discoverable on the global
6:43
knowledge graph. But only those that actually
6:45
have permissions will be allowed to see the contents of it.
6:49
And here you will have stuff pretty deep.
6:52
So they'll be pretty precisely identified.
6:54
Ontologies are going to be very strict.
6:56
You're going to have even down to like identifiers will be
7:00
agreed because it's their partners who they're working
7:02
with. So they have a common
7:04
understanding of the field of even the solution or what
7:07
they're trying to achieve. Like you have importers into the
7:11
US they're, they're, you know, exchanging audit report data
7:14
security audits for overseas factories.
7:16
Like this is a very specialised topic for global trade, hugely
7:20
important because Homeland Security and all those guys can
7:23
kind of check in and see and enable easier imports for, for
7:27
safe products. But it's a pretty tight knit,
7:31
kind of like a, a set up in terms of the, the data
7:34
structures. On the flip side, you also have
7:36
some, some cases where just, you know, schema.org and like very
7:40
open. And it, it can be like, for
7:43
example, the things that we're seeing now with, with, with AI
7:45
agents is they can use the DKG as a collective memory.
7:51
But for them, it's more about how do I, how am I able to map
7:56
out anything that's happening or that I'm interfacing with and
7:59
just publish it on the DG. So like something that's really
8:02
flexible, you know, easy to find.
8:04
That's like a starting point, like a non complicated ontology,
8:07
something that's easy to to interact with.
8:09
And then on on the back of that, you can then build more complex
8:11
stuff. So yeah, you have both.
8:14
And these para Nets can be as narrow or as wide as you want to
8:18
make them. So they go as like I said, use
8:21
case specific or really domain specific as well.
8:26
And for those enterprises, what is the value prop of having
8:32
their knowledge graphs on the DKG if it is completely private
8:37
and it's all for them? Yeah, you know, like it's so for
8:42
example, if you and I are partners and you, we would
8:48
benefit from having more transparency because we can
8:51
optimise our processes better. But The thing is, if you ask me
8:55
to share all of my data with you in order for us to do that, it's
8:57
going to be like, man, you know, like I don't feel that
9:00
comfortable to really do that. Like I'd be much better if I can
9:04
have my knowledge graph. You can have your knowledge
9:05
graph and then let's use these DKG on a need to know basis.
9:09
You had a problem with the product that provided.
9:12
Shoot the query. My system will build, build it
9:15
up and then it will give it back to you.
9:16
You'll get a response. That's response won't be just a
9:19
statement. It'll be a statement with
9:20
verifiability proofs. So at all time you'll be able to
9:23
validate that what you're getting actually has been
9:26
published by our systems at a given time.
9:30
So if we're talking about a product that was produced three
9:33
months ago, you'll see that actually those claims were
9:35
published 3 months ago and I haven't been tampered with it or
9:38
something like that. So you have this verifiability,
9:41
you have this connectivity and and but we can all still have
9:45
our own data in the confines of my system even, you know, like
9:49
we don't have to have I can have one year P you can have another
9:52
one and the things works between each other.
9:54
So it's really good for this interoperability challenges that
9:57
that happened there. So regardless that this is
9:59
private data for everyone outside of our network, even
10:03
between the two, we can make stuff happen.
10:06
That's actually really kind of solving a problem for us that
10:10
that before was addressed. And you're able, I, I can see
10:14
that now because you're able to, to have that transparency
10:21
without giving up or showing everything.
10:27
You can just show what needs to happen and you can verify it.
10:32
And it's almost like there's a, it's not a third parties
10:35
verifying that what you're saying is true, but you're not.
10:39
You're also not just saying it and then expecting the other
10:43
person to believe it just because you said it.
10:46
Or you can have like the audit reports are done by third
10:48
parties. So it's that factory working
10:50
with an importer into the US. So Walmart's buying from
10:53
someone, let's say in China. And then there's a third party
10:56
that's doing the the audit and they're entering their claims
11:00
into the DKG. And then Walmart's able to see
11:02
it. But also the Department of
11:03
Homeland, it's not the same level of depth of what they're
11:05
allowed to see. So you can have very flexible
11:09
kind of management of access rights to to what's done base,
11:13
like almost down to like a line, you know, in the Jason, like you
11:16
can say, OK, this is visible to this entity.
11:18
This is visible to that entity. And you can be really, really
11:22
granular in terms of how you set up these use cases.
11:25
But yeah, like they they can be pretty powerful, even 1 to one
11:28
or especially when the group grows, it's it's even more like
11:33
makes more sense. And how is my data getting
11:37
published on to the DKG? Is it just that there's like
11:42
some messages that are going out and it's wrapped in metadata?
11:47
Am I decorating my code with that and then it gets pushed
11:51
out? What does that look like?
11:52
So the the the most popular way of doing it right now is that
11:58
it's pushed out of the system. So, and it's been pushed out
12:01
towards what we what's now kind of used as a D kg edge note.
12:06
So edge nodes are like smaller lightweight nodes that you can
12:09
do. It's like your gateway or a
12:10
modem into the network. It's like a wrapper around your
12:13
knowledge graph. And then you just populate your
12:16
knowledge graph there. And then you could directly also
12:18
within your systems interact obviously with that knowledge
12:21
graph that you hold there. And that's the whole point of
12:23
it. So you don't have to duplicate
12:25
data, but it's still a specialised knowledge graph
12:28
usually for the use case. So let's say your system of, I
12:32
don't know, logistics triggers an event and that event gets
12:35
pushed and populates knowledge graph that you hold there.
12:38
And then that's being added or is is accessed accessible
12:42
through through a VKG. But if you have an existing
12:45
knowledge graph, let's say in the company or on privately or
12:49
as an individual or as an agent like you can just plug that in
12:51
as well because the DKG isn't a database itself.
12:56
We're the idea is to have as many is many, not redundancy,
13:03
but like options, basically freedom to choose the different
13:06
elements that are used in, in the DKG.
13:09
So you should have a choice of multiple databases when it comes
13:12
to knowledge graphs or you should have a chance of multiple
13:15
block chains when it comes to where do you want to store your
13:17
proofs. So that's the whole point is
13:19
that this is really a protocol that allows to stage these
13:23
things together and play as a common denominator, but allow
13:26
everyone to have freedom of, you know, choosing or your LLM or
13:29
your AI model, like it's not prescribed to use any particular
13:34
single thing. So if I'm understanding this
13:37
correctly, it is allowing you to take the reins and decide how
13:45
you want to craft your own little neighbourhood.
13:49
Like you were saying, in that edge node is the the most common
13:53
way of doing it. You put out an edge node.
13:56
That edge node can have a certain database, it can have a
14:00
certain way that everything is done.
14:03
And then others when they want to gather information from that
14:06
edge node, they just get like the quick, OK, here's how and
14:12
what this edge node is is using. Yeah.
14:15
So if we want to get information, it's it's like an
14:18
API almost. It is that OK?
14:20
We're going to, we're going to grab the information from this
14:24
edge node like this. Yeah, and they can be multiple
14:27
in one parent. So like endless, it could be
14:29
endless edge nodes playing under the same rules.
14:32
You know, even 1 edge node could be a part of multiple para Nets.
14:37
If it probably it, it could even work that way, right?
14:39
So Agenote is like your modem and then you're deciding what
14:42
you want to what you want to do with it.
14:43
But yeah, like the, the, the, I guess the most easy to
14:46
understand example would be exactly like we form a para net.
14:50
Let's say you and I partners will do this para net and then
14:53
other partners will be able to come in.
14:54
You have an edge note. I have an edge note.
14:56
And then we can exchange messages between us or with
15:00
others, but we're as part of this one network.
15:02
And then, you know, if we want to do a global query for
15:05
something unrelated to what we're doing as partners, that's
15:08
also fine. Like you can also do that.
15:11
And I didn't quite understand how the different chains get
15:17
looped into that, because you can also have the flexibility on
15:20
which chain you want to use, right?
15:22
Yeah, so basically it's, you know, we started from like
15:26
Knowledge Graph was our kind of background because we wanted to
15:29
do this transparency thing. And then if the thing was that
15:32
we didn't want to be that aggregator.
15:34
So we, we saw, OK, if you can send Bitcoin from one place of
15:37
the world to the other and there's no bank, like, why
15:39
shouldn't we be able to do something similar with this
15:42
thing that we're working on as a centralised company and just
15:44
take ourselves out of the equation?
15:46
And so then we saw that basically block chains aren't
15:50
real. They're shitty databases.
15:52
They're, they're not really good.
15:53
They're very tiny. They're good at what they do,
15:56
like identities, transactions, smart contracts, perfect.
16:00
But they're not what we were looking for in terms of, you
16:03
know, knowledge graph capabilities like that didn't
16:05
exist at all. So we've basically created this
16:09
decentralised network that that should be on top of the
16:12
blockchain there. It should kind of leverage all
16:15
these cool things about blockchain and enable
16:17
decentralisation, but be much more tailored for Knowledge
16:22
Graph capabilities. So that's why you have the BKG
16:25
notes, which are then they can decide which blockchain they
16:30
kind of connect to and they do so by guessing where there's
16:35
going to be more activity basically.
16:36
So if the blockchain is more optimised for a use case, then
16:40
you can assume that that's going to happen there.
16:43
And then the nodes will Discord notes that the the spine of the
16:47
whole global DKG will be able to go there.
16:50
But it's not dependent on any single chain.
16:54
So that's the point. And what happens on chain
16:55
basically is let's say you're now publishing in a knowledge
16:59
asset onto the DKG. Basically two things happen.
17:03
One is the NFT gets created, which is your ownership roof.
17:08
And with this NFT, because you have the NFT, you're the only
17:12
identity that is able to update that knowledge asset.
17:16
So I cannot enforce an update on what you published even if it's
17:20
a public knowledge asset because only you as an NFT owner, you
17:26
are able to modify whatever you've published.
17:28
And the second part that happens is the as soon as something is
17:33
added or published to the DKG, the proofs get generated.
17:36
So they're like a, you know, a Merkel Merkel tree root hash
17:40
basically gets created and then that gets published on to the on
17:43
chain. So that's a super small thing
17:45
and it's inexpensive to do so. And by doing that either at the
17:50
creation or modification update stage, we always know that what
17:56
I'm reading, whether that's public or private data has been
17:59
tampered with. So there hasn't been a man in
18:01
the middle attack. There hasn't been like, you
18:03
know, someone going rogue and just doing a change because
18:06
they're afraid that they've messed something up or whatever.
18:09
If, if the fingerprint and the data doesn't match, you know
18:13
that it's flawed and like, OK, humans, it's pretty tedious to
18:17
freaking go and click through all that shit.
18:18
But if you're an AI or you or if you use AI system, that's super
18:23
easy to just compare fingerprints, check, check,
18:26
check, yes, you know, check mark and you're, you're good to go.
18:29
So yeah, like these are the core things that are happening on
18:33
chain. And because it's such a
18:34
lightweight usage, we can do integrations basically with any
18:38
EVM chain if there's a desire or if there's a, there's a use case
18:43
that that makes sense for it. But we are not tied to any
18:46
single. The DKG is a is a multi chain
18:49
basically integration. So cool.
18:53
So how have you seen folks using this with agents?
18:58
Now I know that we were just talking about that.
19:01
Yeah, I'm, I think we, we were kind of feeling out chatting a
19:05
few, few times before and then we were, you know, discussing
19:09
how, how fast these things can kind of take off.
19:11
And then when, when we're going to be seeing kind of these
19:14
things happen. And I remember just in October
19:16
also, we were talking with some folks and saying, OK, like when
19:19
will we see like a first artificial entity with some
19:22
monetary gains or like an economic, you know, significant
19:27
pool. And you already are seeing,
19:29
right, that right now happening that in, in, in with AI agents,
19:32
which are autonomous or semi autonomous, being launched as
19:38
entities on that have their own social media presence, that have
19:41
their own wallets, that have their own assets sometimes as
19:44
well that they're pop like creating.
19:46
And those assets are reaching some market caps, which are
19:48
pretty reputable. So you have all these things now
19:52
kind of happening in somewhat isolation.
19:56
So The thing is that and depends on the, you know, different
20:00
frameworks that people are using and capabilities that they have
20:03
as as like someone could set up those agents.
20:05
They have different levels of memory.
20:07
Sometimes they're just more or less like neural Nets, some
20:12
chambers created at the Rec Queen and then like that's it.
20:15
Sometimes they have some pretty like rudimentary memory systems.
20:21
But what we preview to the DKG to really have the strength to
20:25
become is a memory system, a collective memory for AI agents
20:32
where they can, you know, with all the kind of enterprise grade
20:37
level of granularity, they're able to keep their findings or
20:42
their interactions either private and then monetize them.
20:48
Like all these things that people would be or humans would
20:50
be much slower at, you know, OK, go transact, sign, transaction,
20:55
whatever. AI agents between themselves can
20:57
be much faster than assessing. Is that a good deal or not?
20:59
And there's much less emotional baggage at, let's say, creating
21:03
a marketplace where data is exchanged between agents just
21:06
because we're like, you know more about this guy.
21:08
Cool. I'll can you share your
21:11
knowledge about this guy so I know how I will interact with
21:14
him on, on X. And like, this can be pretty
21:17
easily achieved between between the two agents, but also they
21:20
can publish publicly, right? So let's say there's a, you have
21:23
some agents which are tied to, or they have, their mission is
21:26
really to drive, let's say with this whole decentralised science
21:30
field. You know, a lot of agents are
21:34
built with the intention to create public good knowledge or
21:38
scientific breakthroughs. And, and they would be using
21:41
the, the collective memory as a basically by adding to the
21:45
public, public domain of, of knowledge that's available
21:48
there. So agents, yeah, both for
21:50
individual as well As for shared use cases.
21:53
The DKG could be just a perfect platform basically for them to
21:58
to utilise it for for this purpose.
22:02
Yeah, it's, it reminds me of a conversation that I was having
22:07
two weeks ago with my friend Willem on here.
22:11
And he was talking about how they've created a AISRE agent
22:18
and it's there for incidents. And one of the strongest things
22:23
that he's found for their SRE agents is to create knowledge
22:28
graphs of the systems so that the agents can go through the
22:34
company system and say, all right, well, here's here's this
22:39
and here's the last thing that someone said about this PR or
22:43
here. And, and it can traverse and
22:46
really look for what the root cause of the incident is much
22:50
quicker because of these knowledge graphs.
22:53
And it feels like that having something like that on the DKG
22:59
could be very valuable because then it it's almost like a get
23:05
for your system in a way, and not just your code, but
23:11
everything that is around the code.
23:13
That's a good way of saying it. I'm stealing that.
23:16
Yeah, but it feels like it, right?
23:18
It's feels like. OK, well, it's not just the,
23:22
because what William was saying is there's not just the code,
23:26
there's all the developer conversations that go around the
23:31
code, the PRS or the oh, actually there was a whole
23:35
conversation in Slack where we chose not to do it that way.
23:39
And that might be the whole reason that things failed a few
23:43
weeks later. And so if you're able to look
23:46
through the knowledge graph and find that, then that information
23:50
is invaluable when you're trying to root cause something. 100%
23:54
even outside of your systems, why not, you know, feed in like
23:58
the social media if you're an open source project, right?
24:01
Maybe there's someone who like created the shitpost on Reddit
24:04
and like, that's that's your root cause of failing because
24:07
just you know that you didn't respond to that, But like you
24:09
didn't have these things. How will you do it?
24:11
Like you need an agent to just go around and pick these things
24:14
for you, create a nice knowledge graph for it and then make it
24:17
available for you. So I think those things or just
24:20
plug into maybe a swarm of agents who are already doing it.
24:25
I literally had a conversation on X between and with an agent
24:29
and he struck a deal with me that I'll set up his DKG edge
24:32
note and I'll fund him some tracks so he can, because
24:34
there's a token called track the RAC that you're using to publish
24:38
on the DKG. So if an agent wants to publish,
24:41
he needs to get his hand on on track.
24:42
And he was like, I don't know, just starting.
24:44
I was like, cool, I'll, I'll fund your DKG edge note that
24:46
some tracks so we can start publishing it.
24:48
It was done on X. Like I'm like, you know, well
24:50
actually 3 tweets and we had a deal.
24:53
So yeah, I don't know where this is going, but it's definitely
24:56
interesting. Yeah, it does give a little bit
25:00
more agency to the the agents that are out there.
25:04
And I, I wonder what other kind of use cases that you've seen
25:09
that are, have been interesting. And it doesn't necessarily need
25:12
to be agents. It can just be stuff that people
25:16
are doing on the DKG. I'm, I'm really looking forward
25:19
to the decentralised science. I, I mean, I'm really, you know,
25:24
the whole thing we started the origin trail was to kind of be a
25:27
force for something like for the world to be a bit better place
25:30
than we found it. And so a lot of our decisions
25:33
just go towards, OK, if we have a bit more transparency, for
25:36
sure, it's not going to be a negative think in this sense,
25:38
right. So like we, we try to, we try to
25:40
do things that kind of have a positive impact.
25:43
And I think decentralised science has huge potential in in
25:46
achieving stuff. So you have there when you were
25:50
talking about neighbourhoods, there's this whole D side parent
25:53
where they're trying to Crowdsource or create basically
25:56
incentives for people to submit open source scientific research,
26:01
clinical trials, whatever. You kind of is Open Access and
26:04
you're able to use it to to publish it on a D kg in mod
26:07
graph form and then create a swarm of agents on top, which
26:10
will go. It'll just continuously go at
26:12
trying to create new scientific breakthroughs on the back of the
26:16
existing scientific work. And now it obviously it's a
26:20
purely like this is a really experimental thing, but some
26:23
other stuff could be that's more easily achievable is like you
26:27
bring your own data. You know, you have your own
26:30
edge. Note you input your, let's say,
26:32
sleeping patterns or I don't know, you've done your medical,
26:35
you put in some stuff that you got from there, like from data
26:38
perspective. And then you have a local agent
26:41
instance deployed on your local machine.
26:43
So all of it is done on your, you know, privacy preserving
26:45
datas versus like protected. But you're able to access all
26:49
these scientific knowledge that's available on this perinet
26:54
as a public good. And then your agent is able to
26:57
again, traverse shit tonne of stuff much faster than you ever
27:01
could and come back with something that is actionable
27:06
potentially, but also has sources.
27:10
So, you know, and that's why DKG is so cool is because you always
27:14
have knowledge assets lined there.
27:17
Say, OK, I've I found this here, I found this here, I found this
27:20
here. And then on the back of this,
27:23
what became my context, I then produced either summarization or
27:28
pattern recognition or whatever it is that you, it kind of set
27:32
me out to do. And if you have an agent
27:34
framework that it can do this repeatedly in loops, obviously
27:37
you can, you know, enhance the precision of the, the data that
27:42
it captured. You can or expand it or like, so
27:46
you can do this in, in like more times.
27:48
And then we, we multiply it by 10s of thousands of agents going
27:53
at it day in, day out 20, you know, 247365.
27:57
How much more likely are we to to create some like tangibly
28:02
positive outcomes for let's say society is, you know, medical
28:06
challenges or, or other types of scientific challenges that we
28:09
were having just by having a moon orchestrated effort at it?
28:16
Yeah. Oh, I like it.
28:21
The IT feels a bit like it still is fringe, but you could see a
28:27
world where there's some app built on top of it that
28:32
leverages it, right? And I'm thinking about how I use
28:36
a Woop, and I think that Woop has an API and you could have
28:42
all of my data being yeah. Yeah, we've we've done a we've
28:46
done a prototype with the aura ring.
28:49
So my Co founder, he he wears this ring for one sleeves.
28:52
I wear this one, makes me awaken now.
28:56
Doesn't matter. That's the the wedding you.
28:59
Had the wedding one. Yeah, it's a bit more tough one
29:01
to carry. I like it.
29:02
I love my wife, by the way. Different aura I guess.
29:07
Different aura. Yeah, let's see he has that
29:10
aura, the other one and it it obviously captures all these,
29:14
you know, sleeping data and whatever.
29:16
And you can actually extract the data from it.
29:18
And he's done the exact exercise.
29:21
So we've taken a portion of that Open Access works that were done
29:26
on like neuroscience and then sleeping patterns and whatnot.
29:30
But that wasn't a lot. It was just like 10 papers or
29:32
whatever, but it was just a point like, you know, can we
29:34
make this actionable? And he paired it up with on his
29:39
edge, you know, like, I think it was like running a local
29:42
instance of llama or a llama or something.
29:45
And then like that, that was then looking through the pattern
29:49
of his sleep and it got back saying, yeah, you know what,
29:51
like we saw on Tuesdays and Thursdays or whatever you have
29:56
like there's a pattern of you being doing something wrong.
29:58
Again, like going with this author saying, you know, you
30:02
could kind of do this things to improve your, your sleeping
30:07
patterns or whatever. So it's a very nursery rhyme
30:09
type of a thing, obvious, like almost you could, you know,
30:12
figure some stuff out also manually because just the scale
30:16
is smaller. But like if you're trying to say
30:18
jump from 10 papers to 10,000 and if you're looking from, you
30:23
know, if you can, if you have more contextualised, the more
30:25
complex data that you're introducing, it's no longer that
30:29
easy. But for AI, it's it's so like
30:33
that it works well for that. Yeah, I was thinking about how I
30:37
would want to have something where I can just take a photo of
30:44
what food I'm eating. I I'm also with the whoop.
30:49
It's getting your exercise data, but all of the data that you're
30:54
gathering, the more the better. And if it's out there on the
30:58
edge node and there's agents that are continuously checking
31:03
it and comparing it to different studies and different stuff
31:07
that's coming out, then it can be that ideal world where wow,
31:11
I'm, I'm getting like top notch doctor clinical treatment and I
31:17
don't even have to go to the doctor.
31:19
It knows what is going on in the moment as it's happening.
31:24
Yeah. It's, I don't know how, how it's
31:27
going to be called it's agent as a service or, but a lot of this
31:31
type of repetitive or not even repetitive, but like just the
31:37
work intensive things will be able to get automated and
31:41
automation is going to get automated so that we'll have
31:44
actually a solutions or products or agents that will be allowing
31:49
us to cut through a lot of the yeah, tedious stuff that we had
31:53
to do before and just be that. I was just talking with our
31:56
neighbour. She's a 74 year old lady.
32:01
And like, funny enough, she was, you know, oh, I'm, oh, they
32:05
didn't, I don't know, this artificial intelligence, this
32:07
and that and like crypto is hard for me to grasp.
32:10
But then we just got stuck in around, I don't know, I think it
32:13
was real estate or something like that.
32:14
I said it wouldn't be nice that you could just kind of OK, you
32:17
bought a plot of land. Now you don't know what you
32:19
know, you just say to like a programme.
32:22
Can you please figure out what would be like 10 best designs
32:27
for my house with this prerequisites that I want like
32:29
be it you know like super sustainable.
32:32
I want to be. Active and then you have this
32:35
programme that goes cheques the terrain, find similar terrains,
32:39
find similarly well performing assets that were real estate and
32:43
then gives you this as an output.
32:45
Now wouldn't and then OK, but you also can have robots that
32:48
will build that out for you. But the robots, they don't have
32:51
bank accounts. Wouldn't it be easier if they
32:53
could just have wallets where like digital money could be
32:56
transacted much more easily and they could recompensate each
33:00
other for the data that they use, But you can still spend it
33:02
the way you see fit, whether that's crypto or Fiat or
33:05
whatever makes you feel small safe.
33:07
But the end of the matter is that as a, as humans, we can
33:10
have much like much more, much higher quality of life and a lot
33:15
of the things that maybe were like limited before it could,
33:19
could become less limited. And I think if we, but we need
33:22
to do it in the right way, right?
33:24
So, because there's also these rogue alternatives that maybe
33:26
aren't as, as happy a path. That's what we're talking about
33:29
right now. And I, I, I would like to
33:32
believe where I think I, well, I would like to believe, I
33:36
wholeheartedly believe that the decentralised knowledge graph is
33:39
1 component that really makes kind of these happy pets more
33:42
likely versus they do the other one.
33:45
Yeah, yeah. The transparency is huge on
33:50
that. And it it is one of those things
33:54
that helps. It's like sunlight's the best
33:59
disinfectant. Yeah.
34:01
Yeah, no, for sure. I I think it's and because also
34:04
for your use cases, like, you know, OK, like I'll, I'll get
34:06
the best clinical trial. I think the best clinical
34:09
advice, but like, will you trust it just blindly?
34:13
Probably not. Like you want to see this wasn't
34:15
done by some, you know, like off the beat track published journal
34:22
with, you know, semi conspiracy theory thing going there, right.
34:26
So like you want to make sure that there's no bad data in your
34:30
systems when you're doing it. And if you find it just like,
34:32
can you please not like? Yeah, yeah, exactly.
34:39
And so it, it makes me think like, are people interacting?
34:46
So with the data and the knowledge graphs, when someone
34:48
sets up a neighbourhood, are you seeing that there's money being
34:54
transacted because they've set up a great knowledge graph and
34:57
maybe that's valuable for folks. And so others want to come and
35:01
grab it or others want to get access to it so they pay for it.
35:06
But right now is it just all open?
35:08
It's a good question. I think it's, it's more the
35:12
commercial terms currently still are off chain mostly.
35:17
So it's, it's still let's say in between the those that are
35:21
working together and then setting up this use case that
35:23
they, they, they define the commercial terms up until now.
35:27
I feel the big breakthrough could happen with agents because
35:32
again, like less emotional baggage, more straight up, more
35:35
understanding of the value of data.
35:39
Like for them it just be like buying, you know, breath for
35:42
humans. It's a yeah, I need data.
35:44
Of course I can, you know, I'll also sell mine.
35:46
And so like this, this understanding might agents might
35:49
be a bit ahead of than than what we've seen so far, where
35:53
commercial agreements were usually ahead of technology and
35:57
then the technology was there to kind of enhance it or maybe
36:00
provide a new Ave, but it was more there to kind of, yeah,
36:04
keep, keep up with what the commercial arrangements were in
36:07
the back. So this might, this might
36:10
change. And, and like also the paranoids
36:13
were only introduced fairly recently.
36:17
Before it was just, they weren't really a notion.
36:22
It was just want like V6 even before had like the beginning
36:27
was just one big knowledge graph.
36:28
And then this they were done, but they were done with
36:31
implementations. So they weren't like as a, as a
36:34
primitive to be used in the, in the whole network.
36:36
So Paron, it's now actually the D side is going to be one of the
36:39
1st. And then there's coming like you
36:41
have this buzz economy, which is going to be around the different
36:44
data streams and and the different social network social
36:49
how to say it like figuring out signals for either stocks,
36:52
crypto, governments, whatever sports.
36:56
So like you'll have these different parent it's popping up
36:59
now. So I think that the yeah, a lot
37:01
of also monetization is is likely to appear within this
37:08
period now. So I think it's a good time to
37:10
kind of it's still fairly early to to get involved in either
37:14
thinking what's a good parent? It might be how we, how can you
37:16
monetize it like just using all these primitives that are
37:19
available and then yeah, crafting maybe some Nobel
37:24
business models, which also it seems there's more acceptance to
37:27
them in today than it was maybe 3 years ago or something like
37:31
that. Yeah, how do you go about
37:36
replacing old or bad information?
37:40
That's a good question. Yeah.
37:43
It it's, it's one that we were kind of also even from the
37:47
beginning, we were banging our heads a little bit because it
37:49
was, you know, you had you, I mean, some of the use cases we
37:54
had didn't really have like long term valuable data.
37:57
It was more super valuable now, but then maybe for statistical
38:03
purposes. Yeah.
38:04
But when the product is like no longer used or it's consumed
38:09
like food, you know, it, it wasn't necessarily like that,
38:13
that the origin of of that consumed food wasn't necessarily
38:16
anymore that critical after a year has passed, for example.
38:20
So that was something that we we had as a as a challenge from the
38:25
even the first iteration. So what we've said is that we
38:27
want to kind of create like expiry date for any knowledge
38:33
assets that's there that you publish.
38:36
And that expiry date means that the network is no longer
38:39
incentivized to keep that knowledge assets for you.
38:44
So the core notes are they, they won't get compensated anymore
38:48
for holding that data. So they might keep it, of
38:51
course, like there's no guarantee that's going to be
38:53
gone, but they, they, they are no longer compensated for it.
38:56
So it's likely that they will not.
38:59
And because also when it grows, like then they don't want a
39:03
longer to kind of spend their space to, to keep something that
39:05
they're not paid for. So that's, that's one element is
39:10
that you're as, when you publish, you decide how long you
39:13
want to keep something. Again, whatever is relevant for
39:15
you. And then on the flip side, you
39:19
also have a an update feature. So if you want to like change a
39:26
a value or something like that, you can constantly change that
39:29
and just be like, OK, I mean, you could even keep make a
39:31
knowledge assets. That's a point or two somewhere.
39:34
Now, obviously you won't have them the verifiability of
39:37
whatever is behind it, but let's say that you want to have
39:42
reading of an IoT device at a given time or a price feed or
39:45
something like that. So you could have that be
39:49
discoverable, you know, knowledge graph by creating a
39:52
pointer as a knowledge asset saying, OK, if you're interested
39:55
about temperature in this place, shoot an API call there and then
40:00
you will retrieve automatically as a part of the knowledge we
40:03
have that value and you will be able to use it as an in in your
40:07
solution. In what Any way you feel that
40:09
makes sense? For these types of strings, you
40:12
can also have it constantly basically updating, but but
40:17
yeah, like there's there is element of the data is tends to
40:22
stick around. So yeah, it's, it's, you need to
40:25
have kind of as, as, as you design the system of how you're
40:31
publishing something. It's good to keep in mind that
40:34
whatever is the latest state should be always updated in in
40:37
the knowledge graph and in pushed up so that you don't have
40:40
like flopped information. I know that can also become a
40:46
headache with working with like rag systems.
40:52
If you want to update whatever it is document to make sure that
40:57
the the chunks are the most up to date chunks that you're
41:01
giving the LLM or that you have the proper information to give
41:10
to the LLM and not outdated information.
41:13
And so I was just thinking about how you can make sure that you
41:20
have that proper information and the old information you don't
41:25
necessarily want gets filtered out or it falls to the back, or
41:29
you can ideally just get rid of it, like delete it.
41:33
I mean, you, you can with an update, right?
41:34
So an update feature will be like you'll because you'll
41:38
compensate then for deleting literally, you know, you could
41:41
say, OK, I'll delete this and OK, because it's anyways, it's
41:45
such small. So they'll just they'll they'll
41:47
update it with deletion. And then it's going to be an
41:50
empty, basically a knowledge asset thing in in space there.
41:54
Or you'll do you'll you'll do kind of like an update to with
42:00
new state and then the old state is going to get replaced with a
42:03
new state. And that's kind of how you're
42:06
replacing it. But yeah, for rag, it's it, it,
42:10
it might happen also that with. So what we're doing here
42:14
actually is graph rack. It's not just rack.
42:16
So it's not necessarily just documents in terms of blobs of
42:19
text, but it's, they're contextualised in the graph
42:22
form. So like you're already having a
42:25
little bit more context. And then cool thing is that it,
42:30
it's nice if you set up updates because they, the ripple, the
42:33
positive effects, they ripple throughout the, the network
42:37
because if there's an entity, the metrios and there's an
42:40
update, update by me, it, it doesn't have to be done by you,
42:44
but also they'll, we'll have this connectivity happen.
42:49
So, yeah, like it's the the graph rack is, is, is nice that
42:54
there's much more context, much more precise.
42:56
We've done some tests as well there that like compared to
42:58
obviously just, you know, having text is, it's OK, it's better
43:02
than nothing. But in the graph form, you will
43:04
have much more, much more context than actually, you can
43:07
use LMS to query the knowledge graphs.
43:11
They're, they're pretty good at writing out those Sparkle
43:13
queries, like the that's it's pretty performant.
43:17
They've been trained in a lot of sparkle, I guess.
43:19
And like they, they'll be, especially if you give it a bit
43:22
more context or if you give it a template of what it should
43:25
produce, Like it'll be mapping out that initial prompt to a
43:29
relevant query, filling out that's template.
43:33
No problem. Like super, super good.
43:35
And then you can, yeah, just retrieve whatever you need and
43:39
thought, well, OK, since we are now in the rag rabbit hole, let
43:43
me entertain a little bit more is that you can actually do a
43:48
decentralised graph rack. So it's, it's not just over your
43:54
system, but if you do it across the global graph, you actually
43:58
go into anything you have access to.
44:00
So it's like a Federated query that goes down.
44:04
And I mean in in the future when we were talking about like agent
44:09
swarms, you could even do a decentralised, literally a
44:14
decentralised graph RAG where you're triggering other AI
44:18
systems to perform rag over whatever they have.
44:21
So think specialised agents for multiple fields.
44:26
But my I don't know, I can't come up with like a complex
44:28
question all that would trigger in multiple fields at the same
44:32
time. But my agent would then go and
44:35
it would you know decentralised way federate questions to
44:38
multiple agents underneath where each of those would provide a
44:43
graph rag response from their knowledge base and their refined
44:47
model output. Feed it back to my generalised
44:51
model that I started my question with and then I will receive a
44:56
response that's kind of well normalised for whatever my
45:00
initial prompt was. So you have like a whole
45:03
orchestra of knowledge bases and agents or elements or other
45:10
models base anything Gen AI. We kind of figure it like it's
45:13
left hand, right hand, side brain.
45:15
So I just think of agent as the two things together, but you'll
45:18
be able to trigger them like in a, in a Federated way and just
45:23
receive a response. That's yeah, very precise in
45:27
that sense, with all the provenance information
45:30
available, of course. Now if I am if I have my own
45:34
node, I am responsible for how that node is hosted and
45:40
everything that happens there. Or can I send it to you to have
45:45
you host it? What does that even look like?
45:47
Yeah, it's so. So the way you would think is
45:51
that if you want to have your note is basically you're,
45:56
you're, you're cognizant of you wanting to keep your data within
46:01
your environment. So right now it's mostly, let's
46:07
say enterprise is very, very, very adamant to that like they
46:11
want to have it within their environment.
46:13
They hosted their data, their, you know, OPSEC policies and
46:18
everything all pen tests, whether you can think of is
46:22
there not what touches it cool for someone who's like a more
46:27
even for agents, for example, you could do a gateway note.
46:31
So gateway note is like a you know how when you send the
46:34
transaction blockchain, you also don't hold like you don't run
46:37
your etherium or your Bitcoin or your SO on a note.
46:39
You just, you use basically a service that's an RPC service
46:43
and then you sign your transaction with your wallet,
46:45
you append your gas tokens to it and then you send it through
46:50
this gateway. And that same way can be done
46:52
for the DKG. You just find the note, the
46:55
gateway note, someone that's hosting it and you push your
46:58
publish through there. The thing that will happen is
47:01
that that will be it's, it's good for public transactions.
47:03
So whatever you want to publicly put on the public knowledge
47:06
graph. So it's not going to be hosted.
47:09
There's going to be companies that they're doing hosted
47:11
services as well. So you'll be able to kind of
47:14
place your trust to a company so that they do a good job keeping
47:18
your data safe, but it's not exposed to everyone.
47:21
But then it's up to every user to decide whether they trust AB
47:25
or C, you know, but it it's also, it's also possible.
47:30
Yeah, I think those would be like the three, the three
47:34
different variations of it. But idea would be that edge note
47:37
is so light that you can have it, you know, on your phones or
47:40
or laptops or wherever so that it's becomes kind of like a non
47:47
annoying thing that's just there.
47:50
If you need to use it, you can have it there.
47:52
And it's not like a tedious task to kind of run your note.
47:56
It's just an app like you open it, it's there, your data is
47:59
there, it's protected by how you want to protect it.
48:02
And then you want to add your LLM, you want to use public or
48:05
private instance. Again, you have that choice to
48:08
tip and then it just works. Yeah, since it feels like the
48:12
scale is and can be so big and especially what you are trying
48:19
to get to right is the knowledge graph of everything.
48:26
Yeah. How are you thinking about
48:27
speed? Yeah, it, it's like it, like a
48:33
lot of things, it comes down to in a way a good design and
48:39
separation of concerns. So you just mentioned in the
48:42
beginning that we released the version 8.
48:44
So that's yeah, it's, it's something that really amped up
48:47
the scale because we've done some, well, blockchain's
48:51
advanced in some ways, so that's some small part of it.
48:54
But we've also redesigned the the system where because of the
48:59
size of the network, we were able to move towards something
49:02
that's called like a random sampling.
49:04
So we no longer needs to validate every single knowledge
49:07
assets, we can just randomly sample.
49:12
And because of the size, there's no way for let's say a rogue, a
49:15
rogue player in, in the network to be able to of like just try
49:20
to trigger the system and delete parts of it.
49:22
They'll have to persist it. So with that, we've increased,
49:25
we've increased the scale of what we can do by, by like over
49:29
1000 X. So now we are literally like at
49:33
a Internet scale size because even when the knowledge graph
49:38
grows, if it gets too big for the subset of nodes that's
49:42
there, we can split it in two O ways and then the the two
49:45
networks go any then one network again gets too big that can get
49:48
separated. So you have like an ever growing
49:50
basically sharding type of exercise that can happen right
49:56
now when it comes to speed, that's where it's a little bit
49:59
of a separation of concerns because it's like depends on
50:02
what you're trying to do. You can think of it more like as
50:05
a get and search. If you're doing a search where
50:09
you want to do a global search of everything, it takes time,
50:13
but that's just physics. Like it will always take time to
50:17
do it. You can't until we have like,
50:22
you know, multiple processes running at the same time in, in,
50:25
in the, in the quantum, the variations, it'll be hard to
50:30
have everything at any given second.
50:33
So until quantum will have to go this way, and that's going to be
50:37
a bit like it's, it'll take time.
50:38
It's not because you're still traversing.
50:40
It's not like you're going to endless tables.
50:43
So it's still much faster, let's say, than having a relational
50:47
database of everything and available there.
50:50
So you'll have performance there, but what a lot of times
50:53
is you're actually going to forget.
50:55
So you actually know where you're wanting to ask something.
50:58
So it's, it's a pretty precise query that you're shooting,
51:02
let's say to your edge node or to a subset of edge nodes to
51:05
paranoid. And here the speed is, is good.
51:09
Like here is very fast. It's like this.
51:10
This can be production level speeds, deployments without
51:16
because again, it's reversing. It's much, much faster compared
51:19
to having you know something, even if it's decentralised, it's
51:23
still fast. The fastest.
51:25
If you really want to have like the top top top speed, you what
51:29
you do is you do gets and then you place them in your edge note
51:33
and then you just read from your database.
51:35
So you you trust yourself. Once you read it, you know that
51:39
it has recipe. Then you trust yourself.
51:40
You keep it there. And then whatever your
51:42
application is, you just have it interact with your local
51:46
knowledge graph and you'll you'll just embellish your
51:50
knowledge graph with others on a whenever you want to.
51:55
But then like the actual reads happens from the edge node that
51:58
you run. Dude it's so awesome.
52:01
Where can people get started and how can they get started with
52:04
the DKG? There is a Boston community
52:09
around and we'd be definitely kind of keen to to invite
52:13
everyone to check out our Githubs.
52:15
Everything's open source, so you can go in and have a have a
52:18
check out there, you know, set up and match notes, see how it
52:22
works. I connected to to some Gen AIS
52:27
put the two brain sites together.
52:29
If you're setting up an agent as well, like I will be releasing
52:32
some frameworks plug insurance there.
52:34
So the intention is to support a lot of the popular stuff there
52:38
so that you can have it an easier way to to deploy it.
52:41
And of course, this chord would be the channel to to come in and
52:44
hang out. We're we're there daily.
52:46
Our devs are as well. So if there's any, there's a
52:48
Boston community that will help you out if, if, if you have any
52:51
questions. But also our team members are,
52:54
are present. So yeah, we're, we're happy to,
52:56
to have a chat and to build some cool stuff together.
52:59
Man, there's a lot of initiatives already there, but
53:02
I'm so, so keen to learn more from what people are building
53:05
and then how we could make, yeah, fight it a good fight.
53:09
I guess the other it's it's a fun time.
53:12
Cool.