Latest / The Adults In The AI Room / Sean Martin on Practical & Scalable Knowledge Graphs - Episode #9

Transcript

Sean Martin on Practical & Scalable Knowledge Graphs - Episode #9 · The Adults In The AI Room

• 0:07
  SPEAKER_00: Here we go with the leader to the practicality with technical and practical AI in the enterprise.
• 0:13
  SPEAKER_00: As we move from the current state to the future of cognitive assistance and an edit AI in all aspects of our personal and professional lives.
• 0:22
  SPEAKER_00: Thanks so much for listening.
• 0:23
  SPEAKER_00: If this is your first time, we invite you to subscribe for you listening to the podcast.
• 0:28
  SPEAKER_00: And now here are your host, Seth Early and Chris Featherstone.
• 0:36
  SPEAKER_02: Well, let's uh let's go ahead and get started.
• 0:39
  SPEAKER_02: Uh welcome to today's podcast.
• 0:42
  SPEAKER_02: My name is Seth Early.
• 0:44
  SPEAKER_05: I'm Chris Featherstone.
• 0:45
  SPEAKER_02: And uh nice to uh catch up with you, Chris.
• 0:48
  SPEAKER_02: How's uh how's your week been?
• 0:50
  SPEAKER_02: How are you handling uh uh the weather?
• 0:52
  SPEAKER_02: Well, you don't have the weather issues that we have up here in the in the 30 inches of snow.
• 1:00
  SPEAKER_05: Yeah, it's sunny and 20.
• 1:02
  SPEAKER_05: Sunny and 20, okay.
• 1:04
  SPEAKER_02: Have you gotten much snow down there?
• 1:06
  SPEAKER_05: No, we we had um, you know, we get we got dumped on, and then well, it's crazy because we live in the mountains, right?
• 1:12
  SPEAKER_05: So we had an early snow in October, then we had nothing, and then it heated, it warmed up, and then we had another big dump over the holidays, creates kind of what they say is like a marbling effect of the snow underneath the base layers, which is absolutely dangerous for avalanche and slide, right?
• 1:31
  SPEAKER_05: So everyone's on high alert that every everything's gonna slide, but it's been super, super cold over the last little bit.
• 1:38
  SPEAKER_05: And so, yeah, 20, but it feels like it's forward.
• 1:41
  SPEAKER_05: Yeah, it's a dry cold system.
• 1:44
  SPEAKER_02: It's dry cold.
• 1:45
  SPEAKER_02: You don't have the wind going.
• 1:46
  SPEAKER_02: Yeah, so um, before we get started, I do want to thank our sponsors, uh, Simpler Media.
• 1:52
  SPEAKER_02: And you may be familiar with CMS Wire, uh, one of their uh flagship uh sites, the Marketing AI Institute, who has a conference and lots of education, uh educational events and webinars and podcasts and so on, and of course, uh early information science.
• 2:07
  SPEAKER_02: Uh today, our guest is an award-winning technologist who has been on the leading edge of internet technology innovation since the early 90s.
• 2:16
  SPEAKER_02: He's a super smart guy who's written uh a number of uh publications, he has patents uh on his um resume and that he's been awarded.
• 2:26
  SPEAKER_02: He's co-author of an e-book called The Rise of the Knowledge Graph, which we'll start talking about today.
• 2:33
  SPEAKER_02: Please welcome Cambridge Semantics founder and CTO, Sean Martin.
• 2:37
  SPEAKER_01: Uh hi, everybody.
• 2:38
  SPEAKER_01: Uh very happy to be here and uh thank you both for having me.
• 2:41
  SPEAKER_01: It's uh really good to be talking to you.
• 2:44
  SPEAKER_02: Terrific, terrific.
• 2:45
  SPEAKER_02: So, Sean, why don't you uh give us a little bit about your background?
• 2:49
  SPEAKER_02: You're uh I I notice I I know you're you're based in Boston, but you don't have a Boston accent.
• 2:54
  SPEAKER_01: No, I have a pretty mixed-up accent.
• 2:56
  SPEAKER_01: So I I I was born in Johannesburg, so I'm uh originally from South Africa, and I grew up on the eastern, if you think about Africa, the bottom right-hand side of Africa on the on the east coat capes uh is uh city called Port Elizabeth, which is where I I went.
• 3:12
  SPEAKER_01: I really that's hometown for me.
• 3:14
  SPEAKER_01: But I left there when I was 17 and and moved to the UK, where I uh I went to university um up in Edinburgh in Scotland.
• 3:21
  SPEAKER_01: So that's partly where the accent comes from.
• 3:24
  SPEAKER_01: And um, and it was there that I started actually while while studying had an internship with uh IBM.
• 3:30
  SPEAKER_01: Uh and after after completing my degree, joined IBM um uh and ended up working for them for about 15 years, initially first couple of years uh um in um in southern England, um uh supporting what was a a brand new operating system on a brand new uh power uh power uh power chip technology, the AIX operating system.
• 3:53
  SPEAKER_01: Um and I I learned to be a comms engineer.
• 3:56
  SPEAKER_01: Um but I kind of got tired of that after a year or so and moved to uh to um to London, um, where I spent uh a year working with a very interesting group that was called the Early Support uh group.
• 4:10
  SPEAKER_01: And and why it was so interesting is because it had people from all the research labs kind of on assignments in this one uh office, and and essentially they were supporting all the products, the new products coming out of the labs.
• 4:22
  SPEAKER_01: And so I got exposed to all of IBM's uh labs and products in one place, which is really uh really unusual.
• 4:29
  SPEAKER_01: Anyway, it was the dawn of the internet, um, and I built some um early um internet applications for IBM.
• 4:36
  SPEAKER_01: I was really on the leading edge of that.
• 4:38
  SPEAKER_01: Um, and um and that got me invited to go to the US and actually um I got kind of got, I don't know, I don't know if you'd like to say kidnapped, but uh I was I was grabbed by a guy who um who had figured out that some was stealing a march on IBM uh for the Olympics in um I think it was in Oslo, and um and I built a an app kind of a the first one of the first of a kind web applications um and content management systems uh for IBM Europe.
• 5:12
  SPEAKER_01: Um and they thought, hey, why don't we we put the Olympics on the internet uh for Atlanta, which was coming up uh in 96.
• 5:20
  SPEAKER_01: So this was in 95, and there was an agreement reached um whereby I I would go to um uh White Plains uh in New York, where they had a data center and um they'd stolen some dropped off the back uh SP1 parallel computers, um uh which I was pretty familiar with, the AIX machines, and um, and we and the agreement was with the the UK IBM um that if we did the Olympic, sorry, if we did Wimbledon and put Wimbledon online in '95, we could I could stay and then do the Olympics in '96.
• 5:52
  SPEAKER_01: So we not only did Wimbledon that year, and there was the first ever kind of sports scoring system live online, but then we did the US Open.
• 6:00
  SPEAKER_01: Um, and we did, in fact, we did the entire Grand Slam golf and tennis.
• 6:05
  SPEAKER_01: Um, so I did both uh so we we would send people with to trailers on at the at the various stadium and we'd hook into the scoring systems.
• 6:15
  SPEAKER_01: Same with the golf, and we put all of that stuff online, um, all as a sort of a warm-up to uh to the Olympics um in '96, uh, which was um about as close as I ever wanted to sports, I would say.
• 6:27
  SPEAKER_01: After after the Olympics, it was a hell of a it was a six months with literally not a day off and almost no sleep.
• 6:34
  SPEAKER_01: Um I was done with uh these big scale, big scale websites and and so on.
• 6:39
  SPEAKER_01: So and then um I moved to Boston after that.
• 6:42
  SPEAKER_01: And and um at IBM, that team that had put the Olympics and all those systems together really was the beginning of IBM's e-business.
• 6:50
  SPEAKER_01: Um and um uh and so I'd got enough credit to to build another team up in um what was the IBM the year or so before it bought the um bought Lotus.
• 7:03
  SPEAKER_01: Um and so I uh someone gave me an office and sort of drafted me into the into the Lotus payroll with a zero or one dollar salary or something.
• 7:12
  SPEAKER_01: So I look like a Lotus employee and I built a new team there, uh really a Skunkworks that that lasted about 10 or 12 years.
• 7:19
  SPEAKER_02: What did you do with Lotus?
• 7:21
  SPEAKER_01: So um there were quite a few things, but we basically built rebuilt quite a few of the people who'd worked with me on the um the big scoring systems and the and the um the these big websites.
• 7:32
  SPEAKER_01: We also did the tickets for um for Atlanta.
• 7:35
  SPEAKER_01: It was the first kind of one of the first big online ticket applications.
• 7:40
  SPEAKER_01: Um still have nightmares about that one.
• 7:44
  SPEAKER_01: Um uh but a lot of those, quite a few of those guys uh moved up to Boston and joined me there.
• 7:50
  SPEAKER_01: And um, and we did a whole series of of really applied research technologies.
• 7:55
  SPEAKER_01: Um uh one big one, a five-year one, was a was a kind of a JavaScript runtime uh that would run on Windows and Linux and on very early smartphones.
• 8:07
  SPEAKER_01: But IBM, I was told, didn't think JavaScript had much future.
• 8:11
  SPEAKER_01: It wasn't really a serious language.
• 8:12
  SPEAKER_01: So eventually I had to give that up or more or less be fired.
• 8:16
  SPEAKER_01: Uh and and that's around the time I got into semantics, around 2000.
• 8:20
  SPEAKER_02: Did you work with uh Lotus Discovery Server at all?
• 8:23
  SPEAKER_01: No, I didn't really work much with the Lotus products.
• 8:25
  SPEAKER_01: We were a little really this little IBM Island.
• 8:28
  SPEAKER_01: Um what well one big thing we did was we started the IBM Extreme Blue um uh uh intern program, which I was amazed to find is still running 22 years later.
• 8:38
  SPEAKER_01: Um, and in fact, quite a few of the people, my colleagues, now came through that program.
• 8:44
  SPEAKER_01: So so it was it was uh kicked off in in in Lotus, Cambridge in in I think '99 or 2000.
• 8:51
  SPEAKER_02: So and and tell me when you said you got into semantics at that time.
• 8:55
  SPEAKER_02: What did that mean uh back then?
• 8:57
  SPEAKER_01: So so I mentioned that we'd been working on this technology for um essentially building applications using JavaScript and driving, it's a bit like uh the Adobe Air stuff or the Apple, um, these toolkits where you can essentially drive native functions on phones and and and laptops and so on, but use HTML and JavaScript to drive them.
• 9:18
  SPEAKER_01: And we'd built a really elaborate infrastructure for that.
• 9:21
  SPEAKER_01: We had, I don't know, internally within IBM with something like 25,000 users for all sorts of productivity apps.
• 9:27
  SPEAKER_01: And um, but there was there were a suite of apps.
• 9:30
  SPEAKER_01: So we had things like an instant messaging front end and a thing that hooked into the um the the LD app system that our group had set up for IBM and um and there were various informational applications.
• 9:43
  SPEAKER_01: But what we really wanted to do was was have these different we'd actually put a bus into this runtime that you could use to create these apps.
• 9:52
  SPEAKER_01: So you could send messages between these these sort of independent processes that we're running.
• 9:57
  SPEAKER_01: And what we were looking for is some way to move data between them that was more sophisticated than just um, you know, at that time it was XML schema, really, how you could exchange data.
• 10:08
  SPEAKER_01: It was kind of the early days of XML, and but that was kind of too brittle.
• 10:12
  SPEAKER_01: So we were looking for a way to bring new apps online that could hook into the bus and pick out messages that made sense to them, um, without it without having to have agreed all that up front and essentially have these apps really be hard-coded to work with each other.
• 10:27
  SPEAKER_01: So that was the kind of the what we were looking for.
• 10:30
  SPEAKER_01: And and um, and then around the same time, I got dragged into a a really interesting project over at uh Mass General, uh, where a gentleman um uh actually a neurosurgeon was was involved with the uh the National Cancer Institute.
• 10:44
  SPEAKER_01: And he he was um he was looking for help for building a cancer computing model repository.
• 10:52
  SPEAKER_01: Um and as we began to understand what his problem was, it was kind of the same problem, which is how do we how do we share data and integrate data, um, particularly in a when things are moving very fast, which it was in those days.
• 11:05
  SPEAKER_01: They were doing um sort of all sorts of they were part of a program where there were maybe 15 centers around the world that were all doing different kinds of cancer model, but we're talking about in-silico modeling, so simulation.
• 11:16
  SPEAKER_01: Um and and he needed help trying to sort of figure out how to integrate these models or even create a model repository to be the beginning of how to integrate these models.
• 11:27
  SPEAKER_01: Um, and it was the same problem.
• 11:29
  SPEAKER_01: And we we kind of came up with some prototypes around one of the guys who was working with me at the time um headed off to um back to um MIT to just to finish his his PhD.
• 11:42
  SPEAKER_01: And um he happened to be in the uh uh I think he went to the AI AI lab, which is right next to CSAL, which is where Tim Berners-Lee is, and he came back in because it just down the road from from where Lotus was in those days.
• 11:56
  SPEAKER_01: Um he came back once to visit and he said, Hey, I found this this interchange format, it's this thing called RDF, and there's this this there's a standard called Owl, and and we could use that.
• 12:07
  SPEAKER_01: What do you think?
• 12:07
  SPEAKER_01: And that was the beginning of of the semantic journey, which might still am on 20 something years later.
• 12:13
  SPEAKER_05: I don't know about you though, you know, um Sean, but I did something similar, and it's oh it's almost, as you talk about it, almost giving me facial ticks as well, just thinking about the pain.
• 12:26
  SPEAKER_05: I think I we had that we had to translate a power builder app into a Microsoft stack, which was painful anyway.
• 12:32
  SPEAKER_05: But part of that too is the rem is remembering those days of of putting a semantic layer in because the computing power wasn't able to handle it, right?
• 12:42
  SPEAKER_05: And you know, in order to retrieve data that fast and or create, let's say, a configuration engine or something like that, we had to have our read layer, our write layer, and almost a that that meta layer there that the apps could read from fast enough because the systems just weren't set up architecture-wise to actually handle that kind of stuff.
• 13:02
  SPEAKER_01: Yeah, I'm not I'm not nodding vigorously here.
• 13:05
  SPEAKER_01: So you're actually describing my last 20 years, which is uh um which is how to get over those scale issues.
• 13:12
  SPEAKER_01: Um essentially we adopted a graph model that is you know, at the time, you know, we're talking 2001.
• 13:19
  SPEAKER_01: There was yeah, it was Jenna when we were hooking up Jenna to IBM's DB2 graph on relational, it was absolutely awful.
• 13:26
  SPEAKER_01: You get a thousand, you know, 10,000 a million facts, and you're done, you know, and that doesn't really get you very far in terms of practical uses of data, right?
• 13:37
  SPEAKER_01: So for the last two decades have been spent essentially figuring out how to how to do this at scales where it's actually meaningful.
• 13:46
  SPEAKER_01: And and of course, you know, things have moved, you know, Moore's law, cloud computing, parallel computing, very fast networking.
• 13:52
  SPEAKER_01: There's a pile of stuff that's come along underneath the what uh what a collect my colleagues calls the the arc of underlying technology that has been in our favor, yeah.
• 14:02
  SPEAKER_01: Um and uh, but of course, there's also been a lot of software.
• 14:05
  SPEAKER_01: I mean, you're describing also the pain of not having the software, yeah.
• 14:10
  SPEAKER_01: It's not just the hardware that can do it, but actually having software that that is far more um flexible than the sort of rigid ways that we were doing things previously with APIs and very uh um inflexible structures was part of the problem that we were trying to overcome, right?
• 14:26
  SPEAKER_01: This yeah, you just don't know what data looks like.
• 14:29
  SPEAKER_01: And and if you could if you did, then it would be easy, you'd hard code it.
• 14:32
  SPEAKER_01: But if it keeps changing on you, how do you build it systems that are so flexible that uh they can take account of that and not worry about it, not break down?
• 14:40
  SPEAKER_01: And the second you do that, it's essentially at least one layer of indirection you're adding, and probably more if you're doing knowledge representation, and suddenly you're paying a really terrible price in terms of uh performance.
• 14:52
  SPEAKER_01: So, how how to skin that chats and um there are many ways, it turns out, in retrospect.
• 14:58
  SPEAKER_05: That's we actually we actually took our our data from just a relational system because to your point, the software layers, there was also no no native connectivity.
• 15:07
  SPEAKER_05: So, whether there's an object connection or anything like that to actually extract data easily to write simple queries.
• 15:14
  SPEAKER_05: So we, you know, and I'll I'll attribute this to my to my good friend Corey, who was the brainchild behind taking it out and putting it into, let's say, an XSL layer and then reading from that XSL layer because that was the most extensible thing.
• 15:26
  SPEAKER_05: But that was before there was actually any standards for any of that stuff, too.
• 15:30
  SPEAKER_05: And so, which meant it was custom, which also meant it was a support nightmare if something went to hell in a handbasket and you got that one little semicolon out of whack or something, or you didn't have a closing bracket and everything blew up, you know, as you were building simple things, and that's where the facial picks came in.
• 15:47
  SPEAKER_01: I I think I think I think um a lot of people um have had that same experience.
• 15:53
  SPEAKER_01: I think um anybody who used try to use knowledge graph type technology and semantics sort of circuit 20 2000, it actually probably uh as recently as 2015, maybe, um, probably has similar facial text, the number of projects that have created, uh you know, ridiculous number, the number of people had bad experiences.
• 16:13
  SPEAKER_01: So it seems like a really good idea, but when you actually tackle it, you find there's no software.
• 16:18
  SPEAKER_01: Well, the software there is is kind of research level quality.
• 16:21
  SPEAKER_01: Yeah, um, fundamentally it doesn't scale.
• 16:23
  SPEAKER_01: Um, you can't find databases at scale.
• 16:26
  SPEAKER_01: Every part of it you end up doing yourself.
• 16:28
  SPEAKER_01: And in fact, there you go.
• 16:29
  SPEAKER_01: That's my 20 years.
• 16:30
  SPEAKER_01: I've actually had to build the entire stack, or me and my friends, of course, not just me.
• 16:35
  SPEAKER_01: Yeah, so we we we built the database, we built the middleware, we built the BI tools, we built the ETL tools, we built uh every element of a stack because there wasn't anything out there that um actually anticipated uh these flexible knowledge graphy structures and knowledge representations at every layer of the stack.
• 16:56
  SPEAKER_01: So so we had to start start from scratch, and obviously you don't get it right the first time, and it's taken many many iterations.
• 17:04
  SPEAKER_01: Uh, I think we're at eight iter iteration eight, maybe, or a major iteration eight, something like that, over this period.
• 17:10
  SPEAKER_01: So Cambridge Semantics was really um when IBM kind of uh around 2007, I uh IBM was having a doing a lot of cutting, and our group kind of was no one was clear what could understand really what we were doing.
• 17:23
  SPEAKER_01: And so it looked like we were going to be absorbed into back into product groups and things, and and a bunch of us decided that we were much too interested in the technology that we were working on and and and the potential for it.
• 17:34
  SPEAKER_01: So we um so we left and uh and and I I founded Cambridge Semantics, and that's been the vehicle for this technology, which we now have we believe solved all the significant issues and are selling are selling it quite successfully.
• 17:50
  SPEAKER_02: So tell us tell us a little bit about the problems that uh you solve.
• 17:55
  SPEAKER_02: Uh give us an example of a business business problem uh that you guys have uh inserted this year.
• 18:01
  SPEAKER_02: Sure.
• 18:01
  SPEAKER_01: So so so we're a little unique in that we're using knowledge graphs to really um to solve the nuts and bolts of data integration.
• 18:10
  SPEAKER_01: We're not just an endpoint of data that has been integrated, and then you load it into a graph to do graph algorithms and so on.
• 18:18
  SPEAKER_01: We actually use the graph um to to integrate data in graph.
• 18:22
  SPEAKER_01: So we're we have a essentially uh uh the the data engine we have, uh we call it Anzo Graph.
• 18:27
  SPEAKER_01: It's a um it's an MPP data warehouse.
• 18:30
  SPEAKER_01: It's actually the the guys who wrote it were the same people who wrote Parixel, which is the same code that's in Redshift, but that's a relational OLAP um system.
• 18:40
  SPEAKER_01: So it's a it's a data warehouse system.
• 18:41
  SPEAKER_01: Those systems are very different in terms of their performance characteristics to regular, you know, most databases which are transactionally oriented.
• 18:49
  SPEAKER_01: So the business problems we're solving are all around ingesting data as rapidly as possible because these big parallel systems can, you know, by big parallel system, I mean it it's a system that runs as a single logical database, but it's on many servers that are all identical, all doing the same thing at the same time.
• 19:07
  SPEAKER_01: And um, so we slurp in data from any kind of data source as rapidly as we can, and it we can't it can be very fast.
• 19:16
  SPEAKER_01: Uh I think we're doing something like three or four million triples a second for people per node per server node, which is really quite extraordinary.
• 19:24
  SPEAKER_01: Um, and um, and then we allow our users to run transformational queries, just like you do with pl SQL.
• 19:31
  SPEAKER_01: You know, that's been a hell of a successful pattern for data integration for, you know, in in warehouse data integration for decades.
• 19:39
  SPEAKER_01: We allow you to write graph queries to clean up your data and to um um and to knit it together and and represent it in models that the user will will find more compelling.
• 19:50
  SPEAKER_01: So the use cases uh seth to answer your question eventually, things like um monitoring for insider trading.
• 19:58
  SPEAKER_01: Um we're we're where you've got all sorts of disparate data coming in, watch lists and where people are sitting, and call logs and emails and all that stuff, and and building a kind of a comprehensive view of that that an analyst can look to make sure nothing nothing naughty is happening.
• 20:15
  SPEAKER_01: Uh, be an example, or or um uh we're in at the FDA where they have a a very big data lake where all the clinical trial data comes in.
• 20:25
  SPEAKER_01: Yeah, um, and um and they want to build these, uh, you know, the FDA is a large organization with many departments, and and for them to get a single view of a of all the aspects of a drug mean mean pulling in data from these different departments, which are really very siloed.
• 20:42
  SPEAKER_01: So creating these a single view of a drug is an example, um, uh with you know, in in all of its many many facets, the the the trials, the patents, the the label, the yeah, I'm not an expert.
• 20:54
  SPEAKER_01: I'm I'm not I'm not I'm the engineering side, not so much the applied side.
• 20:58
  SPEAKER_01: But that's an example.
• 20:59
  SPEAKER_01: Um another example is uh in in pharma companies um dealing with um adverse event reports.
• 21:05
  SPEAKER_01: So where where something goes wrong with a drug that's on the market, uh big pharma companies are on the hook to report on that to the FDA, particularly if it's serious very quickly, otherwise they face huge fines.
• 21:16
  SPEAKER_01: These reports can come in in faxes and in that system VARS, which has had a bit of bad publicity in the last year or two, or in doctor's reports or phone in, all that stuff has to be triaged and kind of and and if you're a big farmer with a popular drug, you could have you know literally quarter of a million of these reports coming in a year that you have to sift through to understand how serious they are.
• 21:39
  SPEAKER_01: And then, of course, you're looking in these reports to understand you know what's the medical history, what other drugs have been taken at the same time.
• 21:46
  SPEAKER_01: And so there is really quite complex data that you're trying to sift through and understand how uh you know what to make of these things.
• 21:53
  SPEAKER_01: So that's another example.
• 21:54
  SPEAKER_01: Um, let me think some more.
• 21:56
  SPEAKER_01: One that's really up and coming is in the automotive area.
• 22:01
  SPEAKER_01: Where keeping track of um of what lines of production can make what parts across many factories, an example, or integrating um data from multiple PLM systems could be another one.
• 22:16
  SPEAKER_01: So really it's it's it's where there is um data integration problems where it's not going to fit into a nice simple relational model.
• 22:25
  SPEAKER_01: You've got too many different types of things.
• 22:28
  SPEAKER_01: We call that sort of an entity explosion.
• 22:31
  SPEAKER_01: There's often complex relationships between the entities, maybe subclassing relationships.
• 22:35
  SPEAKER_01: These are all much too difficult to practically do um in relational.
• 22:40
  SPEAKER_01: And you end up making very dumb models and you're very limited by how much you can of reality you can represent.
• 22:46
  SPEAKER_01: The knowledge graphs really get over that um very well.
• 22:50
  SPEAKER_01: Um, because you can, uh as I'm sure Seth is very well aware, you can you can you can model you know more or less without limit.
• 22:58
  SPEAKER_01: And um and and that and and more, you can limit you can model at an abstraction level that takes you away from the low-level um artifacts of indices and and um keys and things like that, which just get in the way of human understanding of data.
• 23:12
  SPEAKER_01: So um, so it really special and unstructured data, of course.
• 23:16
  SPEAKER_01: Anything, you know, with the rise of the machine learning, being able to get more and more uh structure, more sorry, more and more information from documents um more cheaply and more um accurately, all that data is of, you know, creates another explosion of types of things you find in the docs.
• 23:33
  SPEAKER_01: You want to tie that data to more structured sources.
• 23:37
  SPEAKER_01: So any kind of application that that anticipates a wide variety of heterogeneous data that you've got to tie together.
• 23:44
  SPEAKER_01: And and where it's really going is is is um is what we call the data fabric, where you're trying to get a view across all of the data in a business and how it how it ties together, um, and find ways to describe that and and and and and and be able to deliver dimensions or views or data products on that um for different business purposes.
• 24:06
  SPEAKER_01: Um so that that's where this where we're going with this technology.
• 24:09
  SPEAKER_05: I think Sean, that's the probably the most the most uh you know valuable thing, like just in terms of the last segment of structure to unstructured could be people have no idea what the hell to do with it.
• 24:23
  SPEAKER_05: One, like you said, structure to structure is is uh is pretty straightforward.
• 24:28
  SPEAKER_05: However, you just have to figure out, you know, from that perspective on the on the outcome, you know, what is it that you want to match together?
• 24:37
  SPEAKER_05: But there's you know, it may be some you know, some interesting matches, but you're still going to get some pretty relevant data quickly.
• 24:44
  SPEAKER_05: When you start trying to figure out where all that unstructured data is, so like you said, if you can even index it because you don't know what it is, then you bring it in and nothing is structured.
• 24:56
  SPEAKER_05: And what we're talking about for everybody out there, it's really just the documents and the knowledge doc, you know, knowledge workers that are creating all the time, whether that be health spreadsheets are pretty easy, but you know, Word documents or um or you know, PDFs or whatever, where it's kind of hard to get access to what's contained within those that information.
• 25:15
  SPEAKER_05: And then how do you pull that in and actually apply it to a structure that can actually be made relevant to the rest of the organization?
• 25:21
  SPEAKER_05: Right.
• 25:22
  SPEAKER_05: And so, you know, you absolutely have to have some type of uh umbrella that gives it the ability to understand what's even out there so that your business systems can make sense of it and also that your people can make sense of it, and you can actually then derive out more value, you know, you know, for the organization.
• 25:41
  SPEAKER_01: I think that's key because how many documents are you know, millions of documents within they say it's it's they say it's something like 80% of the data is unstructured, you know, compared to structured.
• 25:52
  SPEAKER_01: Yeah, although I think there's an awful lot of structured data that's kind of hidden away in X's databases and spreadsheets and things too.
• 25:59
  SPEAKER_05: But uh you think of like just in terms of so how do I take and this this is not necessarily a you know um uh well it is is a pretty comma.
• 26:09
  SPEAKER_05: I have all this sensor data and the sensor exhaust coming off all my machines.
• 26:14
  SPEAKER_05: I'm a car or I'm manufacturing or whatever.
• 26:16
  SPEAKER_05: How do I take that sensor data that's coming off in you know, you know, maybe multiple times a second, and then apply that to a business value outcome tied to a Word document that was created by let's say some product manager by trying to drive some outcome.
• 26:29
  SPEAKER_05: Exactly whatever, right?
• 26:31
  SPEAKER_05: I mean, you name it.
• 26:33
  SPEAKER_05: That's the problem that we're running into.
• 26:35
  SPEAKER_01: That's that's you're you're exactly outlining the kinds of things we're doing.
• 26:40
  SPEAKER_01: Yeah, and uh what I tell you is really quite exciting is that on the forefront of the machine learning stuff with these big um big word models, um, they're starting to really get close to the point where they can populate ontological structures, knowledge representation structures, from directly from text.
• 26:58
  SPEAKER_01: And we're starting to, you know, that's one side of the coin, and the other side of the coin is is uh natural language query, you know, and they're they're very similar.
• 27:06
  SPEAKER_01: Yeah, well, there's many similarities in those problems.
• 27:09
  SPEAKER_01: Yeah, um, that is really exciting because it it moves the ball quite significantly from where we are using what I'd call uh we call them annotators, which are essentially our are all banner of could be as crude as a regular expression, it could be some statistical based thing or dictionary or who knows what the approach is.
• 27:28
  SPEAKER_01: But an annotator, they tend to be relatively inflexible and are varying degrees of accuracy, and but they know what they're finding, kind of and um, but they're they're you know, so so but but what you're gonna find then is really um um what those annotators are are are uh are are built to find, and they're and they're really very specialized.
• 27:51
  SPEAKER_01: And uh previously building these things and getting them right was very expensive, but but it looks like people like NVIDIA and it and Google and others are now starting to put these these um AI models out that will completely change the game there, I think, in terms of um of making it possible to to decide what you want to get out of a text.
• 28:13
  SPEAKER_01: Um, you know, maybe you'd be reading company reports or um you know, or I don't know, maybe it's drug labels, who who knows what the uh what what the uh and and then populate those structures fairly directly.
• 28:24
  SPEAKER_01: Um I think that is gonna really improve over the next few years.
• 28:28
  SPEAKER_01: Um and um and so where the analytics that are currently we're doing with unstructured data is more about, you know, there's obviously sentiment is one that everyone knows about.
• 28:37
  SPEAKER_01: But but you know, finding occurrences of products in in the literature, SKUs, phone numbers of people, emails, thing, ways to identify people, but it's it's still relatively kind of crude.
• 28:48
  SPEAKER_01: Um, you can create these nice meshes, you know, nice knowledge graph which ties this document to this data set and and so on, so you can see that, but it's it's not quite the same as literally extracting the data itself out and using it as if it's structured data.
• 29:04
  SPEAKER_01: And I think we're going there faster and faster, which is pretty exciting.
• 29:08
  SPEAKER_01: I've been reading some interesting papers of people on the forefront that actually at IBM um research.
• 29:13
  SPEAKER_01: There's a whole team that's published some interesting work on that, and there's uh there's plenty of other things going on out in the research community.
• 29:19
  SPEAKER_01: So I'm looking forward to that because we just we're we're downstream of that, right?
• 29:22
  SPEAKER_01: It makes our we do just so much more valuable.
• 29:25
  SPEAKER_05: Oh, yeah.
• 29:26
  SPEAKER_05: And well, I mean, I I love the fact too that the natural language query ideas are based off of some maths, you know, like um sentiment-based models, right?
• 29:35
  SPEAKER_05: Almost in terms of like almost like how we used to do with chatbots and everything else, like, hey, here's some sentiments.
• 29:41
  SPEAKER_05: Those deep learning models are gonna come up with the most relevant information based off of what I'm trying to ask it.
• 29:47
  SPEAKER_05: Now, what we're saying is I'm gonna see what my data looks like today.
• 29:53
  SPEAKER_05: Well, that model was also turning a ton last night because I added you know more data and more data points to it.
• 30:00
  SPEAKER_05: So now, not only am I gonna bring that back with some level of accuracy, but I'm also gonna get be able to ask for reinforcement learning to make sure that what I'm asking is actually becoming more accurate so the model gets better over time, right?
• 30:12
  SPEAKER_05: And I think the I mean, we're just barely starting to see these types of techniques come into play, not really for broad business use yet, but you know, these things like what you're doing with the semantic layers are essential, not critical, like like critical, but they're absolutely essential.
• 30:30
  SPEAKER_05: They're cable stakes now in order to get this done.
• 30:32
  SPEAKER_05: If you want these capabilities, you have to have these ontology architectures that make sense in order to help the models understand what's even in the data or associated with the data.
• 30:44
  SPEAKER_01: Well, that's actually the other thing is generating ontologies automatically from a body of text.
• 30:49
  SPEAKER_01: Now that is very useful too, right?
• 30:51
  SPEAKER_04: Yeah.
• 30:51
  SPEAKER_01: And then obviously you're going to probably want to curate that and and you know, all the normal stuff.
• 30:55
  SPEAKER_01: But but uh yeah, and the other elements, uh and you kind of alluded to to it the in fact you alluded to a number of things that are all interesting.
• 31:02
  SPEAKER_01: One is context, right?
• 31:04
  SPEAKER_01: You know, I that there's extra data coming.
• 31:07
  SPEAKER_01: I'm talking to this thing, and I'm um, but there's extra data coming in.
• 31:11
  SPEAKER_01: The context of that extra data means that you know there's a concept of more accurate, more up to date, for example.
• 31:18
  SPEAKER_01: That so context is part of that.
• 31:20
  SPEAKER_01: Um, and and the other thing I was thinking of, of course, is that knowledge graphs currently underlie the the most popular sort of consumer versions of of uh um natural language.
• 31:32
  SPEAKER_01: I don't know if it's query or search or some kind of hybrid, but yeah, you know, I'm I'm not gonna say the word because it'll it'll chime in here, right?
• 31:39
  SPEAKER_01: But everyone knows little the little hockey pucks uh from uh you know, those are big knowledge graphs underneath them that provide that context.
• 31:47
  SPEAKER_02: Um uh so it's people shall remain nameless at the moment.
• 31:51
  SPEAKER_05: Well the problem is the problem is as soon as you as soon as you say their name, then they think that they're first class citizens of the conversation.
• 31:59
  SPEAKER_05: Yeah, you don't necessarily want to have them.
• 32:02
  SPEAKER_02: Yeah, Sean, tell us a little bit about the book that um that you recently published.
• 32:08
  SPEAKER_01: It's it's very short.
• 32:09
  SPEAKER_01: Uh it's it's only 80, 80, 79 or 80 pages, but it's really uh it's a good read.
• 32:15
  SPEAKER_02: Go ahead.
• 32:16
  SPEAKER_01: Well, it's not a it's the butler didn't do it, but um uh it it's a good read if you're interested in in kind of how to get into knowledge graphs.
• 32:25
  SPEAKER_01: It's kind of what what are they?
• 32:27
  SPEAKER_01: Um you know, the you know, trying to break it down.
• 32:30
  SPEAKER_01: What's a graph?
• 32:31
  SPEAKER_01: What's so nice about graphs?
• 32:33
  SPEAKER_01: Uh what's knowledge?
• 32:34
  SPEAKER_01: Uh, how do we represent knowledge and try to break that down as simply as possible in something uh short and sweet and and really sort of give people uh the beginnings of what they need to start applying this technology at home?
• 32:47
  SPEAKER_01: I think there's I mean we're seeing a huge upswing in in in um in interest in graph and knowledge graph, um, you know, commercially.
• 32:56
  SPEAKER_01: And um, and this book is simply to help people who are sort of knowledge graph curious get their feet wet with something very accessible, lots of you know, diagrams and pictures and simple examples, and then use cases so that they can not only educate themselves, but then be able to talk to their colleagues and so on.
• 33:12
  SPEAKER_01: So that's that was the intention of the book.
• 33:15
  SPEAKER_02: Um and just just for the listeners, just for listeners, there is a link uh in the show notes to download that.
• 33:22
  SPEAKER_02: So be sure to take a look at that.
• 33:24
  SPEAKER_02: And I do want to once again thank our sponsors, Simpler Media, uh, which is uh which uh you might be familiar with CMS Wire uh website, uh the Marketing AI Institute, and uh early information science.
• 33:39
  SPEAKER_02: So uh let's keep going.
• 33:41
  SPEAKER_02: And uh we're talking uh with Sean Martin, who is the founder of Cambridge Semantics, and we're talking about uh AI, knowledge graphs, ontology.
• 33:51
  SPEAKER_02: We haven't talked about ontologies.
• 33:52
  SPEAKER_02: Do you want to uh give your definition of ontology and then how that relates to knowledge graph for our audience?
• 33:58
  SPEAKER_02: Oh sure.
• 33:59
  SPEAKER_01: So um so an ontology is a way of describing knowledge.
• 34:06
  SPEAKER_01: You know, if if a knowledge graph is is um a graph of interconnected data, um, so business data, for it to be useful, enterprise data, some you know, data that's of use to some community, um, you need a way of essentially describing what is in there.
• 34:25
  SPEAKER_01: Um in a in relational world, you you have something which's a little bit more low-level because it describes artifacts of storage as well as the the model, but it's a schema.
• 34:35
  SPEAKER_01: Um and an ontology really is a kind of a more abstract stream schema for representing knowledge.
• 34:42
  SPEAKER_01: And ideally, it's uh it's using a standard so that when you you can pass this this uh this data representation around to uh other parts of your community or other communities and essentially establish a shared understanding of what's um of what you expect to find in uh in the data.
• 35:03
  SPEAKER_01: So it's it's really a it's a template for defining what's in your data and that is and and described at an abstraction level that is really makes it knowledge.
• 35:12
  SPEAKER_01: Um, and so um so you to make a decent knowledge graph, you need an ontology to describe the business data you're gonna find in there.
• 35:21
  SPEAKER_01: And ideally, that ontology speaks to the the domain that the data describes in ways that the people who will consume that data understand it.
• 35:30
  SPEAKER_01: So it should be in plain business English that that the users of the data will will understand.
• 35:34
  SPEAKER_01: So if you're a you know a radiologist, you'll have one set of you know way of describing things versus someone who who deals in maybe research around drug safety or or and so on.
• 35:44
  SPEAKER_01: So so it's it's really a way to and and the the advantage of using a standard is well, there are actually many, but one is that you get to use lots of people's software, um, and um uh because they're all able to read on the same standard, so you can you can move this data easily um out of wherever it started to lots of different systems without any transformation.
• 36:07
  SPEAKER_01: So it's it's essentially the the ontology can accompany the data and describe it.
• 36:12
  SPEAKER_01: Um and one of the wonderful things about this is it kind of future-proofs your data for when the application that created it no longer is around, but it makes that data you know self-describing so that software can come along and say what's this and adjust itself around that data, read that data in, and then start to do useful things with it.
• 36:32
  SPEAKER_01: So that's what ontology is.
• 36:34
  SPEAKER_01: The knowledge graph needs the ontology um to describe the describe the data.
• 36:39
  SPEAKER_02: Yeah, I think of the ontology as the knowledge scaffolding, yeah.
• 36:46
  SPEAKER_05: But it you know, it also organizing principles, it also gives it a a definition point too to write reference back where you know, Sean, you brought up a really good point that that can and may need to change over time, and or there's no nothing that says you can't have multiples, right?
• 37:01
  SPEAKER_05: That describe so you know, but it also it gives some um we'll say standardization, not in the in the colloquial sense of like, hey, there's a standardization for graphs, but it gives an organization the ability to base something on a specific um definition type that they can then at least have some semblance of like structure in a world where there's you know, obviously a lot of moving parts and say, okay, well, you know, that's going to be for this, you know, business graph that we're actually utilizing this for these applications.
• 37:33
  SPEAKER_05: I think it's you know, the the one thing that I want to make sure folks know is that it's not a one-size-fits-all, it's very semantic, and that's the beauty of that is it's the semantic nature of what it can be within an organization can be applied in thousands.
• 37:47
  SPEAKER_02: Right, right.
• 37:48
  SPEAKER_02: And and again, the the knowledge graph uh is really the date is the ontology with the data.
• 37:54
  SPEAKER_01: Yeah, exactly.
• 37:55
  SPEAKER_02: Mechanism, right?
• 37:57
  SPEAKER_01: I think the other thing that's cool about these is how you can kind of stack them together, right, and connect them up.
• 38:02
  SPEAKER_01: Um so you can you can, you know, there's a lot of reuse of knowledge here.
• 38:06
  SPEAKER_01: Unlike schema, which tend to be really fixed and static and tied to a particular instance of a particular database, um these ontologies really are are are free and and can even you know either uh by themselves be reused anywhere, but but also parts of them can be reused in other ontologies.
• 38:27
  SPEAKER_01: So you can you can dip in and say someone over there has defined this concept of a drug or a person really well.
• 38:33
  SPEAKER_01: I just want to use that here in my definition of it.
• 38:36
  SPEAKER_01: So that that's a particularly good thing.
• 38:38
  SPEAKER_01: And it and and and it's it lends itself to this notion of the fabric, right?
• 38:42
  SPEAKER_01: Where yeah, where you where you're you're starting describing certain important, let's say, business concepts, your products and customers and whatever.
• 38:50
  SPEAKER_01: Um, but but as you as you um start to expand your fabric, you're adding these additional concepts and you're creating linkages, the relationships, um, which is why we want to use graph because it's great at holding these connections between different things, like kind of like real life.
• 39:06
  SPEAKER_01: Yeah, and uh and so you can expand out very easily um from your ontology, either by borrowing other people's or creating your own, um, or both, more likely.
• 39:18
  SPEAKER_05: Not to be pedantic, but you could actually have a network of graphs or even a map, a graph that actually has the data that holds pointers to others as well, right?
• 39:27
  SPEAKER_01: I don't I don't think that's being pedantic.
• 39:29
  SPEAKER_01: I think that's pretty accurate, right?
• 39:30
  SPEAKER_01: It's what people want to do.
• 39:32
  SPEAKER_01: It's uh yeah.
• 39:32
  SPEAKER_05: I mean, I didn't you know it's I just didn't want to be like the you know captain obvious, but there's there is that notion though of like we're talking about where you know organizational, you know, there's not I and I'm with you, man.
• 39:44
  SPEAKER_05: I hate you know the notion of let's say you know a hardcore schema that's on a structured environment and it's predefined and it gives you no flexibility, right?
• 39:53
  SPEAKER_05: As opposed to the you know, transform and load and definition after the fact.
• 39:58
  SPEAKER_05: Yep.
• 39:59
  SPEAKER_05: So you can actually have you just need the flexibility.
• 40:01
  SPEAKER_05: And I think people get into this the colloquial well, no, it's you know, this schema against this schema bullshit.
• 40:09
  SPEAKER_05: The ability you've got to have flexibility within your structures because there's it's changing all the time.
• 40:16
  SPEAKER_05: And if you don't, then you're gonna miss and you're gonna have you know gaps in abilities to answer questions that you didn't have before.
• 40:25
  SPEAKER_01: Totally.
• 40:25
  SPEAKER_01: And uh there's another thing we've observed.
• 40:27
  SPEAKER_01: Uh I mean, I'm absolutely seeing what you're saying in spades.
• 40:30
  SPEAKER_01: Another thing though, we've observed is that people when they when they treat ontology as like structure, they kind of kind of create the ideal ontology.
• 40:39
  SPEAKER_01: Then you when you actually go to apply it, you find that that that they haven't got either haven't got the data or to to actually you know to create instances of of that ontology, you know, of data conforming to that ontology, or or they've just missed missed a whole lot of stuff that turns out to be really important in the in their exercise.
• 40:58
  SPEAKER_01: And so what we find is the is you know maintaining that flexibility and essentially modeling the data you have is very important, and then maybe some upper upper ontologies to try and give you kind of um I guess a highway map to the concepts and so on.
• 41:14
  SPEAKER_01: So you get these fairly low-level ontologies that bubble up out of the data, they look a bit you know, schema-like, but you try and abstract them as as much as you can, and then ideally try and connect them to upper ontologies to make them navigable and to help you find data.
• 41:28
  SPEAKER_02: Um Sean, have you seen or worked with uh uh sumo suggested upper merged ontology?
• 41:36
  SPEAKER_01: I you know, I I I I looked, I have looked at it, it's ringing a bell, but I would as I was just saying to Sharon before this uh cost, uh the uh everything everything is is being forgotten here.
• 41:48
  SPEAKER_01: So I have looked at it.
• 41:49
  SPEAKER_01: I've just I don't know if I've ever used it though.
• 41:51
  SPEAKER_02: Yeah, I I was just curious if you bumped into it, uh colleague.
• 41:55
  SPEAKER_01: I have.
• 41:59
  SPEAKER_02: But it's the idea of mapping to everybody.
• 42:01
  SPEAKER_01: That's right.
• 42:02
  SPEAKER_01: It's quite a lot, it's like 45,000 concepts or something.
• 42:04
  SPEAKER_01: I'm having a big bell ringing here.
• 42:07
  SPEAKER_02: Yeah, and tell me where you see the tools and technologies and capabilities going.
• 42:13
  SPEAKER_02: Um, you know, what what's what's on the you know, what what's practical today versus what's aspirational?
• 42:19
  SPEAKER_02: And where do you see okay uh that uh North Star, that uh that ontological knowledge graph on uh uh uh Nirvana.
• 42:31
  SPEAKER_02: Go ahead.
• 42:32
  SPEAKER_02: Do you want to answer?
• 42:32
  SPEAKER_05: No, no, I was gonna say, Sean, you can tell us as much of your roadmap as you want, but uh just kidding, go ahead.
• 42:39
  SPEAKER_01: So um so where we are today is we can handle really large knowledge graphs.
• 42:48
  SPEAKER_01: Um, we can build out these clusters on our engine that that let you handle you know tens to hundreds of billions of of facts.
• 42:56
  SPEAKER_01: And that's now a practical reality.
• 42:59
  SPEAKER_01: We have that stuff in production at lots of you know places, and and um and what people are doing with it essentially is is is integrating data, usually through ELT style queries, and then putting and then and and those some of those transformations will be from the these kind of raw, raw schema-oriented ontologies to canonic what we call canonical ontology, but business facing, business user-facing ontology.
• 43:26
  SPEAKER_01: So so creating data products is is absolutely a reality.
• 43:29
  SPEAKER_01: Blending in unstructured data to the extent you have good annotators for that unstructured data, that's also a reality.
• 43:35
  SPEAKER_01: Um, and that gets you pretty far.
• 43:37
  SPEAKER_01: I mean, it's a much better way of integrating data than anything we've seen out there before, right?
• 43:41
  SPEAKER_01: So that that's what I think we are.
• 43:43
  SPEAKER_01: Where I think it it goes now is into some of the things I alluded to earlier, which is as the annotators get better, we're going to get much better with unstructured data.
• 43:50
  SPEAKER_01: I think the natural language query is something people really want to do.
• 43:54
  SPEAKER_01: Um, but I also think we're gonna start seeing in graph um machine learning, um you know a lot of the the the the machine learning community is actually using kind of graph structures to do these neural networks and things i suspect all of that's going to collapse um um either through improved hardware systems from from the you know people like NVIDIA and Intel and um um as well as improved software from from many different companies um and so I think the analytics you're gonna be able to do on the integrated data without pulling it out but actually doing it in place um is going to be very interesting in the I mean here's the thing the the the graphs that we use the graph engine we use are essentially schema less right so we can ingest dirty data without you know just pull a file in turn it into a graph on the fly and then you let you manipulate it but but if you go to the far end and you've now integrated it you call it blending data or rapid data engineering whatever it is you've blended all of this into a nice beautiful knowledge graph now you're gonna want to start to do things that infer links automatically that cluster things automatically um uh that predict parts of the graph automatically but but at the same time you're gonna want to keep track of that which is you know kind of done by the human and is almost certainly well probably right versus that which is inferred and which models are used and which training data and so the graph lends itself to to on the fly restructuring its for it's not actually a schema but the ontologies you can grow as you go so you can you can start to do all sorts of interesting things with metadata because really there is no unlike in you know a warehouse world where your metadata and your your data are generally kept in different systems metadata can be part of the graph and can often be part of the what you need to answer a question.
• 46:05
  SPEAKER_01: Queries to proven the provenance of the data yeah so the exactly so the provenance of data is is huge right so all of that comes in and and and um and similarly when you're going the other way where you're starting to predict data I think we'll be using the graphs to to hold all the provenance of of of the predictions too yeah um so I think that's kind of in the immediate future.
• 46:27
  SPEAKER_01: I think um the other thing that's got my real attention at the moment is that the the technology I'm describing currently is only available to really um the most forward thinking sophisticated environments at this point.
• 46:42
  SPEAKER_01: It's been very expensive to develop this right we've and so our customers were self-selected to be a very you know very big companies who have very um interesting use cases that make it worth their while to be on the leading edge of this I think we've now reaching a point literally in the last year or so and where where this is this technology is ripe to go out in in a much more um democratic expression where it'll become available to to um um more many more types of individuals sure you can always go and download uh you know neo or use one of these it's sort of what I call transactional graphs and you can get it you can load a certain amount of data and so on but you're not really going to be doing at any scale the sort of stuff I've been describing.
• 47:29
  SPEAKER_01: I think we're very soon going to be able to let a lot more people use that technology for whatever they want to do with it at outside of the sort of fortune uh you know outside of the government and the fortune 500 where it's been really a domain and the the tech the the software is getting easy enough to use and uh the people can do it themselves.
• 47:50
  SPEAKER_01: Part of the problem of serving smaller businesses and and groups is really the expense of the services that surround that and the um um and the uh you know it's just you have to you've had to have traditionally really good people to help people to guide them on the way but but there's a certain level of automation and self-service that I think is is is now where it's becoming viable for people to really help themselves.
• 48:13
  SPEAKER_01: So I'm looking forward to that I think that's what's coming in the next year or two is is is a much wider access to this kind of technology and who knows what that's going to let people do.
• 48:22
  SPEAKER_01: You know if it's it's easy to access and it's reasonably cheap um we're gonna see an explosion of these much more sophisticated data based applications.
• 48:32
  SPEAKER_05: So I'm look that that that that's near term uh well we'll definitely see you know obviously like you said we'll see tooling get better right give you more flexibility and stuff and almost it's almost like moving from quants to be uh to business analysts you know what I mean like that transition at the same time you know the those primitives into building blocks into managed services that could just be used and consumed you know from a organizational structure you know still developer based but you know what I mean like there's you know some some pieces because to your point um you know a little moving from that that notion of of having these primitives and moving them into a business application that makes sense there's a lot of people don't necessarily care how the sausage is made they just want the business value out of it.
• 49:19
  SPEAKER_01: Yep but the beauty of this technology and this layer that you're that you've created or that you know that you're really focusing on is if people really understood how much of this is the connective tissue between their data and business value they'd be doing it 100% of the time and that's the tippy point right totally totally agree and and up to now it's been you've really had to select the highest value applications you can barely afford to speculate um uh because it's been really too expensive and and the software has let people down over the years for sure many people have been burned um I'm you know we've been burned ourselves um we've all been burned we've all been burned um but but it once that cost drops down and and there's more automation driving it um it's a it's completely changes the game then you know everyone says data's the new oil but and and of course the big sort of speed hump in the in the way of of of machine learning has been you know getting integrated data into and and of course train training data sets right once you get beyond you know if we can solve those issues and get at it and make it you know significantly cheaper I think um and really we're talking about the tools getting that much better yeah I think is the essence and and and and the cost of it coming down so that there's less uh less expertise needed to get quite get further much further yeah there's well if data is the oil but it's also changing into solar power too and you need a semantic layer between those to bridge the gap right to figure it out so well we're right at uh just about at the top of the hour and uh again want to thank uh simpler media c mess wire marketing institute early information science and of course sharing for uh all of her hard work behind the scenes and uh Sean it's been great having you chris thank you sean any final words in terms of uh advice or takeaways or you know what's the low-hanging fruit if someone was going to go back to their organization and try to make a business case what's what's that killer app or that uh I don't actually begin that process do you have that the the the ebook that you referred to earlier the very last page or so is essentially advice this advice that I'm gonna repeat which is essentially get started now and pick something that is relevant to your business that has drives business value and keep it small but show off the principles of knowledge graph knowledge engineering and so on in that and the success of that will will allow you to um uh to grow it and to do adjacent applications or expand that one yeah success breeds success so uh don't try and boil the ocean don't try and build a beta fabric whole whole cloth do it um as dean alarm one of the co-authors uh said do it a stitch at a time and and uh that would be my advice i don't know what your killer app is and out there in the audience it's going to be relevant to the people who pay you and it's gonna be valuable to them but pick something that you can use as a as a as a a vehicle for this um and that's that that that has worked for me and and um that's why i'm i'm repeating the advice so it all it always it all goes back to gol's law right let's do something simple yep so that we can make something more complex out of simplicity as opposed to make it super overcomplex in the beginning that's not the that's you'll never finish right no no you won't you won't which is probably why you know schema based systems have reached the end of their limit we need all these other great technologies to connect them up so I I think so I really do think so it's um pleasure thank you for having me thank you Sean thank you Chris it's been great and uh talking again soon and uh having you back uh until then thanks everyone good weekend everyone yes you too okay thanks for listening to the Early AI podcast with your host Seth Early at Chris Featherstone we hope you enjoyed the program and took away some ideas that you can start to implement in your organization right now.
• 53:36
  SPEAKER_00: If you're stuck and need help check out www.early.com for information about early information science consulting services and of course check the show notes for links to any resources we mentioned today that's all for now see you next time