Full transcript
0:06Philip, uh, over breakfast the other
0:08day, you were telling me that the graph
0:10databases made by your company Neoforj
0:13were used, uh, by journalists in the
0:15Panama Papers investigation in 2016,
0:18which exposed um, a global tax evasion
0:21scandal that implicated heads of state.
0:24Could you tell me how graph DBs were
0:26used and what they are?
0:28Yeah, what was actually interesting in
0:31the investigation in this data leak was
0:35not, you know, the the legal papers and
0:37documents and who owned this particular
0:40legal entity. Um because the people many
0:43of the people in that database were
0:44actually trying to hide funds. And the
0:47way they hid funds was uh by creating a
0:50chain of entities which owned each other
0:52and then having people who maybe shared
0:55an address or were spouses or were close
0:57business associates um al also offiscate
1:00ownership through um having their family
1:03members own things. And so even if you
1:07had just you know gone deep and analyzed
1:10it in any other way you would have
1:13missed these connections because they
1:15were oftentimes many levels deep like 10
1:17levels deep. Um and so the ICIJ came
1:21across Neo4j as a way to take all the
1:23different entities and the relationships
1:25between them and
1:28um essentially make sense of it and find
1:30the trail. So it's it's sort of like you
1:33know you're you're swimming in all this
1:35noise and all of a sudden you can
1:37extract things and see the connections
1:38and it just snaps and they were uh you
1:42can still go to the website and see some
1:44of the visualizations and the ownership
1:46chains and this had pretty big global
1:48consequences like some prime ministers
1:51and ministers were uh lost their jobs as
1:53part of this for having not disclosed
1:55assets that they owned. Um and uh that's
1:59that's really the heart of it is that
2:02the world shows up as these complex
2:05networks and if you can represent them
2:07as a graph you can suddenly understand
2:09the way things are connected you can
2:11understand the context and you can
2:12understand cause and effect much more
2:14deeply. So there's a whole database uh
2:17market. I guess there's SQL, there are
2:19like more traditional databases. Just to
2:22be clear, you're basically a direct
2:24competitor to those sorts of
2:25technologies. It's uh it's competitive.
2:29It's also complimentary. Um there you
2:32know starting in the early 2000 2010s
2:35there started being a lot of alternate
2:37models to the relational database of
2:39which graph is one but you also have uh
2:42you know uh column family document
2:45databases like MongoDB uh key value like
2:48reddus so um
2:52when you're building systems you
2:53oftentimes need different technologies
2:55to handle different specific kinds of
2:58data analysis so there are certainly
3:00cases where Neoforj has been uh brought
3:02in to replace those technologies. But
3:04more often than not, it's just playing a
3:07role in a larger larger ecosystem of
3:10many technologies.
3:12>> I think yesterday you said that if
3:14people were designing databases from
3:16scratch today, they would probably
3:17design a graph database.
3:20>> What did you mean by that? What I meant
3:21is if you're trying to design a database
3:24to solve the problems that show up in
3:26today's world, well, the kind of data
3:29that you need usually is, you know, some
3:32digital real world system that I need an
3:34agent to understand
3:36and where I actually also need a common
3:39understanding amongst multiple agents
3:41like one need look no further than molt
3:44bot, molt book, open claw, whatever it's
3:47called today. um uh to see the kind of
3:50chaos that can uh occur when you don't
3:53have some um central uh grounding force
3:58both from a reasoning and from a um to
4:02to to kind of fight against the entropy
4:04that otherwise descends. And uh given
4:07that the world tends to show up as
4:09networks whether it's biological
4:11ecological networks of payments,
4:13networks of computers, networks of
4:14people, ideas, spread of disease uh or
4:17as hierarchies which is a specialized
4:19form of network shaped more like a tree,
4:21org chart, asset asset ownership, supply
4:23chain that uh if I want to solve for
4:26those problems, then ideally I would
4:28want a representation that is shaped in
4:32the same way those show up in the real
4:33world, which is a graph. Um and then
4:37likewise modern compute has evolved
4:39greatly from the time relational
4:40databases came into being. So the reason
4:42you have a rigid structure is partly
4:45because they were designed to solve
4:47business process automation and
4:49digitizing paper forms at the end of the
4:51day
4:52but also subject to the constraints of
4:55memory uh spinning disc at the time
4:58where you couldn't just randomly chase
5:00across data to connect the dots to
5:02understand larger systems. you sort of
5:04had to to narrow the aperture. So for
5:07most applications being built today um a
5:11typical application should actually you
5:13should start you know certainly consider
5:16a graph database to start with and in my
5:19estimation probably at least half of
5:21applications uh should start with and
5:24you know use a graph database at its
5:26core.
5:27>> I think that's pretty compelling. I'm a
5:30former engineer. Uh, and the nerd envy
5:32like rejoices to hear that this idea of
5:34graphs being a kind of natural
5:37representation for the world. Um, but
5:39the journalist in me notices that graph
5:42databases currently account for about 2%
5:44of the global database market. Um, why
5:47is that do you think?
5:49>> Exponential curves start slow. You know,
5:52it's like the drop drip drip drip of
5:55water in the stadium and 30 or however
5:57many iterations later when you're
5:59doubling it every time, you know, you it
6:02it it looks like the water is barely,
6:04you know, filling the base of the
6:06stadium and then you blink and it's it's
6:08all full. Um so part of it is this uh
6:12data the average database management
6:14system from the research I've seen is
6:17deployed and then that application and
6:19its database live for about seven years.
6:22Um and then you also have people who are
6:25taught in school to think and learn in a
6:27particular way and it takes some time
6:29for them to make their way um into the
6:31workforce and see things differently.
6:33Um, so the database space is one that's
6:36particularly I think slow to move
6:39relative to certainly consumer tech and
6:41there's a whole spectrum between that.
6:43Um, but what gives me assurance is is
6:48I'll point to two things of of several.
6:51One is
6:53the almost every developer that I've
6:55seen work on on a project with Neo4j
6:59will start to see the world in graphs
7:01and will start to see how oh the world
7:03is networks. What am I doing putting it
7:05into tables? Like it's so much more
7:07natural and easy to work with once I've
7:09learned to think in that in that way and
7:11work in that way. Um and then on the
7:15other end, standards oftentimes define
7:18the next generation of technology. And
7:21after more than 30 years of having just
7:23one ISO standard for database management
7:26systems, namely SQL, um ISO uh around
7:31two years ago, April 2024, released a
7:34graph database
7:36language standard which defines the uh
7:39property graph model and which defines
7:42the language around it uh which is
7:44called SQL. And this is more or less
7:46looks very similar to Neoforj's cipher
7:49language. you can use cipher and have
7:50assurance that that uh is GQ SQL
7:52compatible and um so that you know at
7:56these two extremes developer adoption
7:58and the behavior of a developer once
8:01they've gotten into the tech as well as
8:03um
8:05you know ISO standards define the next
8:08generation of technology. They're not
8:09meant to define something that's just
8:11going to be around for a year or two. um
8:14as well as the fact that AI applications
8:19uh in order to be successful in the
8:20enterprise where you have a higher bar
8:22for success, it appears that you need
8:25something like a knowledge graph for
8:26which the best implementation is a graph
8:28database. And then last but not least,
8:31AI makes it easier to overcome some of
8:34the hurdles that have always ex existed
8:36for people to adopt graphs, namely
8:38learning the language. Now you can just
8:40have agents translate human queries into
8:43cipher or SQL queries. Uh and likewise
8:46creating the graph AI is really good at
8:48taking on that what was previously a
8:50piece of heavy lifting.
8:52>> I want to come back to the AI question
8:54and and we should definitely speak about
8:55that. Um you mentioned exponential
8:58growth being slow at the beginning but
9:00it's always exponential. I mean has your
9:02has Neoforj's growth been exponential so
9:05far?
9:06Yeah, I I joined so I was having a
9:09dinner with a VC a couple nights ago and
9:11we were talking about Neo4j. I've been
9:14at the company around 14 years. Uh we
9:16raised our series A um just over 13
9:21years ago and he kind of sat back and he
9:23said, "Look, it always takes 12 years
9:26before a company is even known by
9:28anyone. Uh and then everyone thinks it's
9:31an overnight success and this thing just
9:34turns around." So, in the time I've been
9:36there, I joined sub million in revenue
9:39and before anyone knew about the
9:41category, let alone the product. Um, now
9:44we're in 84, the Fortune 100. Uh, we we
9:48have um hundreds of startups that I know
9:50of, hundreds of AI startups that I know
9:52of just in the last few months that have
9:54signed up to our startup program. Any
9:56startups, I'd welcome you to to do that.
9:58Um and uh you know and some of the
10:02biggest buzz in AI in the last few weeks
10:04has been around knowledge graphs and
10:05context graphs. So I feel like um and
10:10and you know to answer your question as
10:12far as the growth to date, you know,
10:14we've gone from sub million to uh at the
10:17end of last year we crossed 200 million
10:19in ARR um and we're last valued at two
10:24billion. Uh so uh definitely the growth
10:28curve has been like that.
10:31>> I've managed to go 11 minutes of this
10:33talk without saying the word context
10:35graph uh which is I think the title of
10:37the panel. Can you yeah give a little
10:40more detail on what that is concretely
10:41what a context graph is.
10:43>> Yes. Um so a context very simply is and
10:48we all know what it means you know just
10:50kind of in human language it refers to
10:54um tell me not just facts about this
10:56thing but give me knowledge about things
11:00around it where that thing came from how
11:02that thing behaves and then perhaps how
11:04the things around that behave and the
11:06things around that behave and if I look
11:10at it maybe a bit more rigorously
11:12you could look at the con at graphs at
11:15having several different layers of
11:16context or kinds of context. At the
11:19highest level, um you hear about context
11:23in the world of um semantic layers and
11:26this refers to where does the data come
11:28from. So you have the context for a
11:30piece of data which you know one
11:32definition is well what system did it
11:34come from and when was it last updated
11:36and who updated it. So that's one kind
11:39of context. The other kind of context is
11:41let's take an example in consumer web
11:44and I'm having uh I'm wanting to give a
11:48recommendation to someone. They have
11:49three things in their shopping cart.
11:50What should they put in next? Well, I'll
11:52do a better job of recommending the
11:55right thing if I understand what things
11:57do people usually buy together when
11:59they've already bought these things. So
12:00that's the context around those things.
12:03What has this person bought in the past?
12:04Have they bought these things together?
12:06What have family members in their
12:08household bought? what are things they
12:09might have returned. Okay, let's
12:11probably not recommend that one. Um, so
12:13that's another form of context around
12:15the actual domain of the data. And then
12:18there's yet another form of context
12:20which is uh what um you know what's a
12:25lot of the AI discussion has been around
12:26the last few weeks which is why did why
12:31was a particular decision made
12:33understanding that so that in the future
12:37I can use that information to make
12:39better decisions either by capturing the
12:41rules and executing those rules
12:43deterministically or by informing an LLM
12:47and bringing that context in either as
12:50pre-training or as part of the prompt to
12:54um to make a better uh LLM informed
12:58decision in the case where the reasoning
12:59happens there and what you can do with
13:03this which ultimately is what's valuable
13:07is once I have one or more of these
13:09different kinds of context which by the
13:11way each interrelate with one another
13:14but you can start small And you once you
13:18have it you have essentially a source of
13:20knowledge that machines can read and
13:23execute on and then humans can also
13:26understand. So it becomes a bridge
13:27between human knowledge and AI knowledge
13:32human valition and AI you know whatever
13:34it's going to do. Um, and it can also
13:37become a layer that's used by multiple
13:40agents so that they can coordinate with
13:42each other. So they all aren't just
13:44looking at different parts of the
13:45picture. Like I think about the classic
13:49uh ancient Indian tale of the seven
13:52blind men and the elephant where each
13:53one is touching a different part of the
13:54elephant. It's like this thing's like a
13:56snake and this thing's like a brush and
13:57this thing's like a tree trunk. And this
13:59is what you get with agents that are
14:01each using silo data that's
14:03disconnected. So it can give you that as
14:05well.
14:06Does
14:08this picture of the context graph
14:10fundamentally change the work that
14:12Neoforj is doing? Does it change the
14:14technology that you're building or is it
14:15basically just a new application for the
14:17same technology? It
14:18>> it fundamentally changes the stack that
14:20it participates in. And so to that
14:22degree, then it changes everything at
14:24the connectivity layer in in in runtime,
14:27what systems need to connect with it,
14:29including like having an MCP agent uh
14:31layer and connecting with agents. Um it
14:35uh at the development time layer it
14:38changes um it offers new possibilities
14:41for developing in terms of code
14:43assistance. Uh so we've capitalized on
14:46AI for helping users write queries. Um
14:50it also changes um I'd say the altitude
14:54at which you use the graph. So many
14:57applications historically have been I'm
15:00going to use the graph for one
15:01application or one use case. And then
15:04over time you would end up with a
15:06particular enterprise using you know
15:07having lots of different graphs and then
15:10they might bring them together or
15:11federate queries against them. And here
15:15what I'm seeing a lot more of is the
15:16pattern of let's deliberately go in you
15:19know kind of at a CIO co level and say
15:24we actually need an organized
15:26store of knowledge distinct from data
15:28like you you may have a data lakehouse
15:31somewhere where you've put every bit of
15:32data but that's not suitable for real
15:36time access and shared knowledge nor
15:38does it know anything about the
15:39connections. Typically, it's hard to do
15:41multiple joins, let alone multihop, let
15:44alone transactional.
15:46And so having a system of context that
15:51agents can access and it supports both
15:53applications so transactional uh access
15:57patterns in the database as well as
15:58analytics where you can actually run
16:02graph AI so supervised and unsupervised
16:05learning like you know Google was the
16:07first I think big commercial example of
16:09this of page rank let's prior let's rank
16:13order all the search pages based on how
16:17many inbound links I have from other
16:18pages and then weight those based on the
16:20number of inbound links. So there's a
16:22whole rich set of graph theory and graph
16:25algorithms that can be applied um on the
16:28analytic side to actually enrich your
16:31context based on taking what you know
16:33about the topology of the graph and how
16:35things show up and roles that
16:37individuals nodes or actors play and
16:40feeding that back into your AI system.
16:42So, it it it it's a way to also bridge
16:45all of the kinds of AI that we've yet
16:48invented um alongside Gen AI to make
16:51your Genai applications more powerful.
16:56Here's a totally different question
16:57because we have time for one more. Um my
16:59friends and I sometimes talk about uh
17:02what our post AGI job would be. What we
17:05would do if artificial general
17:07intelligence tomorrow solves all the
17:08problems and we can just sort of do
17:10whatever we like. Uh what would your
17:12post AGI job be?
17:13>> What would my post AGI job? Yeah. Well,
17:16let me first say that I don't feel like
17:19AGI is necessarily helpful as even a
17:23mental model because it it assumes that
17:25there's sort of this one monolith that
17:28will do everything perfectly and give
17:31you perfect answers. And my experience,
17:34you know, particularly in B2B and in the
17:36enterprise is there's not just one thing
17:40or one intelligence that is operating as
17:43part of your AI system or your
17:45application. You actually have a
17:47composite of big models, small models,
17:49each with checks and balances playing
17:51different roles. Of course, knowledge
17:53graphs and other other tools. And so to
17:56ascribe general intelligence to like one
18:00part of that which is the LLM seems
18:02misplaced.
18:04Okay. Having said that, what would I do?
18:06Um
18:08I would apply myself to seeing how we
18:13can not use it as a crutch but using it
18:15as an amplifier of our own human agency
18:19um and our own action and creativity. Um
18:23I was in a
18:25conversation a few weeks ago uh with
18:28some people who worked with foundation
18:30models about, you know, how how do you
18:32get LMS to answer certain kinds of
18:34questions better um to help people um
18:38navigate their lives better? And one of
18:40the things that we realized is you
18:42actually don't want the LM to just hand
18:44someone the answer on a silver platter.
18:46That's super irresponsible because then
18:49we just become auto automatons. We don't
18:52you know that that's uh how many more
18:55steps do you have before we're all
18:57encased in pods in the matrix with you
19:00know some jack in our heads. So um I
19:02would work to uh
19:07to enable AI in such a way that it
19:11supports and enables human potential
19:12however intelligent it may be.
19:15>> Cool. Thank you.
19:19>> Thanks. It's a pleasure.