Knowledge Graphs are key to unlocking the power of AI

[Music] hi everyone and welcome to another episode of architect tomorrow I'm really pleased to be continuing the conversation around the latest advances in machine learning particularly around generative Ai and today we're going to be focusing quite a bit on Knowledge Graph that technology and how that kind of plays into the ecosystem and the architecture that we're looking to build and so I am pleased to be joined by another panel again you'll notice there's three of us there will hopefully before Charles has got it we're stuck in traffic on the way to London the joys of doing in person rather than Zoom recordings let's kick off with some some intros although you'll recognize some faces if you are watching the previous episode on the kind of General risks and challenges around building generative AI powered applications in Enterprises but I have one I'm Chris Booth product owner of machine learning an OS group my primary focus is looking at immersion Technologies creating prototypes and improved Concepts to on to bring value to the bank specifically I specialize in competition Ai and help manage the quora artificial intelligence agent portfolio hi I'm Tony seale and a software developer I've been specializing in doing knowledge graphs for investment banks for the last 10 years or so now and recently then I've been looking to combine that with large language yeah and it's it's been great to have you join us Tony because we've really appreciated the post you've been putting on LinkedIn on topic um so yeah do go and check those out if you haven't already I should say at this point before we go any further as pretty much always architect tomorrow is a personal podcast it's our own personal views the views of the community not those of our current or previous employers last time we were talking about the kind of components and the attributes that we needed in in Enterprise architecture and machine learning we talked about a lot of the risks and the challenges and some of the technical challenges and one of the interesting things um I think I will probably go on to talk about this a little bit Chris is some of the ways we've used knowledge graphs practically to kind of mitigate some of the failings I suppose all the challenges that large language models and geometry AI presents but before we kind of get into all that Tony you're the expert in the space for over 10 years of of knowledge I just wondered if you could perhaps just give us a bit of a you know 101 for those not really you know know come across knowledge graphs before I mean I I did a bit of digging and I saw that Google in 2012 introduced the kind of Google Knowledge Graph which is when you do a Google search you get that sort of nice information box on the right hand side that's powered by Knowledge Graph but perhaps give us the sort of 101 for those that have not come across an orange graph sure well so it may be a good way of thinking about it is it's a different way of structuring the data so we're at the moment we're kind of used to seeing data in tables in relational databases and Excel spreadsheets and like if you're a developer then you're used to seeing in XML files or Json files those are like tree life stretches and the graph structure is a different structure where we basically have nodes and then we connect them with edges and with a Knowledge Graph what we do is we have three elements that let us do that so for instance I could have used the subject and then I can have the eye color as the predicate and then I could have blue as the the value within there and the funny thing is if I then said well here is another subject and you guys are friends and and now we've got a link there and amazingly with that you can basically create a graph uh and then interestingly um with the knowledge graph stuff you can actually use um Uris in order to be those identities so uniform resource in identity that's right yeah you could just call them URLs really because uh actually it makes sense that when you go to one of those URLs it's actually resolved below for HTTP that turns out to be quite important yeah but the interesting thing there then is that you can take your knowledge graph and you can distribute it out over the clouds right now all of your data does not need to be co-located all into one place but this can be a graph of interconnected data that is distributed on the cloud and so is this essentially the kind of semantic web made real uh because there was a lot of clatter about semantic web did that term just sort of fall out of favor and it's quietly kind of been happening behind the scenes yes it's the same thing yeah that's kept all the semantic web is dead or now it's going to be called link data and now it's called knowledge graphs right it just it kind of won't it won't die and because I guess because it's actually a good idea yeah yeah yeah yeah I mean it's a bit like uh with the large language models neural Nets oh they don't work it's you know actually that's a good idea he just needed the computer yeah yeah and so one of the things that caught my eye is when you spoke about Jason LD so Jason linked data um and so I guess is that sort of the Pinnacle now where we've got the sort of semantic website Technologies using link data to kind of represent things that allow us to connect up and make relationships between things in the Enterprise is that Etc where we're at yeah what I find particularly interesting about Jason LD is um there can be a thing with graphs that they're very complicated or you know this you know I understand where I am in the relational database I understand where I am with Jason but graph that's kind of sounds like kind of scary and uh Jason of the really kind of sorts that out for you because it means that you can just put your data in plain or Json and then you can have a couple of attributes at the top to say you know this is the context this is where my kind of schema is going to be and this is what this attribute ID is going to let me put the URLs uh bring those into the picture and then suddenly any developer who is able to write Jason is able to produce fragments as a graph and they have so now over 40 I think it's like 44 of all web pages on the web and you know like this is the point that I always want to have a home to people this is now a widely used technology so right and these aren't this is not like rocket scientists or anything like that the web developers skilled excellent very uh highly rated web Developers who are trying to optimize themselves in the search engine when they're putting their HTML Pages out there they're including in that within there like here's a little bit of data here's a little nugget of data about when you mentioned briefly um like some people will find it a bit more complex or difficult relative to SQL or relational how much of that is just a habit familiarity and how is it like actual it is more complex my personal experience yeah it took it had it took a while to head around the different structuring of the data with the grass and open the edges but once I've done that recently I found I actually managed to Traverse and understand and build graphs quicker than they do as well yeah it's it's I agree with you for me it's actually more intuitive it's closer to how I think you know if I was to go to a whiteboard and try and we were trying to swap an idea then you end up drawing something you've got to kind of force it into this structure yes it's very unnatural as much as much it's a much lower level structure really where we had to worry about performance and obviously you still do need to worry about performance performance matters but basically yeah we can push that down to a lower level what do you think this is the main stick pushed down for marriage again or maybe some people prefer have graft brains so you will have relational brains but it is relationships it's just it's just sort of representing them in a different way right I think relational databases have always been about relationship between things but it's been like I say a bit artificial to think about things in a tablet because that's not how we think about the world they yeah there's the irony there they call it relational database the relationship is not actually made a first class yeah it's true it's just the keys suggestions the keys of them yeah and then yeah yeah the irony um so so yeah I mean look um the reason Chris and I were already came to talk about this is that we've done we've done some work together actually learned a bit of a proof of concept uh if I can get the okay maybe what I'll do is is uh overlay in a minute into the poster overlay some of the demo of what we've done but um I'd be very interested to hear about that yes let's talk about that a little bit Yeah so um we were working with uh with Chris uh and we were kind of doing some joint r d around these uh large angles are great that kind of create conversational content but how how would you use them in a highly regulated sort of sector how could you control them stop them going off the deep end and sort of talking about anything so one of my colleagues in London actually came up with a really nice ontology that represents uh conversation and the objectives of the conversation and so hopefully also we can share that actually to get sort the thumbs up and we'll we'll share that I think it's really I would say it's the finished product but what it does is it essentially has the the it's essentially a graph of a conversation and the outcomes that you want to sort of drive and the sort of data points that you'd like to know in that conversation and so the beautiful thing about the album is we then take that and we look at what's the weakest connection on that Grant as we sort of cycle through and that's then what's used to drive the prompting vlms ask the the user the next question so you know it could be we don't know the basic information about the customer so we need to kind of get some Basics filled in on that graph or it could be we you know the example's actually a financial health check example so we need to know their sort of outgoings or their income so that's the sort of piece of information we need to take and then the the first thing that the demo will show is as the sort of conversations happening the graph is kind of getting built out and the Beautiful Thing of course about the llm is it doesn't have to be precise that you know it's exactly this pounds Shillings and Pence you know it's it'll it'll pick up so smoke talking about you know 500 quid or you know talking about actual conversation like you would to do as a human being but it's managing to extract that and put that information into the graph along alongside the kind of conversation that's happening so that's the proof of concept that I'd love to see so that people sort of see and hopefully it's it's useful um and that that kind of start the whole conversation about what are the things we need to kind of put around the edges of some of these generative Technologies because they're exciting but let's face it UNESCO building so that's very b2c very consumer orientated whether the risk is low of getting something wrong if you're in a highly regulated sort of market then you're going to need to sort of think quite carefully about what this thing is doing for your brand or any other things we talked about in the last episode so yeah it's it's exciting that we it didn't take us actually that long it was a few weeks really to kind of get it's amazing isn't it yeah this piece is the building blocks are there so that's what's exciting for me is that yes there's a lot of hype around generative AI but actually the thing that gets me excited is when we combine this with other technologies that are more mature uh that's when I think we'll start to see some really really interesting things happening in the business space I think we'll see all kinds of crazy stuff let's say in the consumer world you know we've seen lens uh you know we all sort of have a flavor of that but I think when it comes to actual practical you know Enterprise applications you know integrating it with your integration platforms you know knowledge graphs is just a huge control here so what excited me talking about your post the other a while ago I guess probably a couple months ago now um was sort of the idea of using Json LD because then you're putting for your Enterprise data because then potentially you're putting the data in a similar sort of format as because your point earlier the common cruel data set that these large language models are built on is built on lob semantic web data which has Jason LD in it so it kind of made absolute sense when you were saying get your organization's data into the same sort of format I mean yeah can you talk a bit more about sort of that sort of concept yeah and I'd like to drill into some of it because I I made some notes while you were talking sure yeah and you brought up the the over the ontology word so that that I think maybe for the benefit of the yeah yeah the audience you probably should kind of drill into that and I think they'll fit into the Json LD thing a bit as well so um the autology is the kind of like is this they've got a semantic part of the semantic web so there what we're trying to do is we're trying to um extract out the abstract Concepts if you like so um not at the level of the individual data but the kind of schema of the data is the metadata conversation and out on with the Jason LD then that is provided by schema.org so that's where a bunch of big search engines all got together and said well you know okay let's create a ontology a model of the various things that people talk about on the website right you know this is a shop this is I've got this product that must be pretty comprehensive really yes it is it it's big and flat like if you kind of go and speak to a professional ontologist uh you know they will say it's the worst ontology that you know it's not because it's too far yeah if you've ever heard someone say that's a good alternative that also is very true unless the person break the ontology themselves yeah um but uh what it is is widely adopted so that you just absolutely can't um that that you can't deny so um when all of these people are um saying this is a customer this is an airline this is a flight this is a product then they are all referencing back to that same common model that what it is right so that's that's that's powerful right there so then what you've done in your own ways you've developed you're saying an ontology about what a conversation is so this is the agent this is the yeah so that kind of abstract level and an important understanding is really any any and I would I would say every business should be about this now like one of the concepts what are the words that we use to talk about our business let's get those and let's get them into an ontology and then you talk about uh schema.org and the large language models but so what I say here and it's this is a relatively controversial point of view but I I do think there is logic in it as you're pointing out don't re reinvent the wheel kind of use schema.org as your basis why do that well because there's already it's already been widely used yeah so like take your version of customer and the nice thing about ontologies and graphs and stuff is we've got inheritance in there so you can just extend that out to add on your own properties on top of what their customer is I mean one of the key messages that I that I'm trying to spread out at the moment by doing these things the AI Revolution is happening like what that that one I've been talking about this for quite some time I know that a lot of other people have as well there's an exponential curve for this thing it is coming it's now even beginning to bleed down this Consciousness every business needs to be consolidating their data so that's I've said it before that we're basically we look we look on at the tech companies oh my God they're doing all this incredible stuff but each individual organization is like a beggar sitting on top of a gold mine of all of the information that they have collected over all of the years they've got all the experts there who really understand what their domain is but it's in a mess yeah it's it's all over the place it's fraction yeah you can't if not I mean maybe one day we we train an AI that's kind of powerful enough open up the bottle and announce it goes and sorts it out yeah and who knows one day that may happen but I don't it's not a good business strategy so there's a human problem that you still have to go over right even if the AI can restructure your data your human processors and other things need to change to kind of be compatible with that I mean it's interesting to sort of hear you talk about this sort of business but you kind of agree your business ontology or agree your business terminology because it feels like that would solve so many different things in the organization anyway right I mean like Knowledge Management challenge for example I'm talking about widgets you're talking about items or skus it's just silly things like that where Enterprises sort of triple themselves because they've got so big that they can't have that you know um conversation around a table like we're having now to agree terminology you need that set of agreed terms and I think there's actually a bunch of reasons why finally getting on top of data and taking it seriously because I think you're right the difference between the tech startups and the incumbents predominantly the fact that the tech startups have come this green field and they're able to build cleaner data architectures and clean up cloud architectures and insert your architecture here but um that for me is the fundamental difference people get sort of um scared by sort of tech destruction but that fundamentally is just the fact they've come at this with a blank sheet of paper are you aware it might it might be it sounds like a schema is schema to always the leading the way is there any efforts to create one ontology to build them all well basically schema.org is there and then we have at the moment is as close as we have at the moment and it uh a kind of 40 coverage within within the web it would be there and and also that tends to be the one that people are going to in order to like extend say for instance within the financial domain we've got fibo pushes okay the the business ontology for finance yeah yeah but then there's a kind of lightweight offering of that which is included within schema.org so depending how you want to go deeper will have a notion of data set data catalog in it but then if you want to go kind of deeper on that then there's the decal on the top why haven't they merged or wanting to the other sometimes I suppose it's hard to subject to say ones well I think the way that it's going to go is that you will get detailed business aligned ontologies and then you'll kind of get these kind of like wider broader Authority so schema.org is a wider broader ontology and then you can create mappings between that's what we found out is like for specific use case you're going to have to perhaps you guys your ontology starts taking off you create the detailed conversation and yeah I think probably we could do with something like that it sounds like a useful solution for sure yeah no so maybe that's what you end up doing and you go really deep on that but I would say that when you do it have a chat to the schema.org guys do a pull request from there online because because what strikes me about sort of schema orgas I guess it's it's born out of search engine optimization broadly I mean like a massively simplifying I appreciate it probably skipping over a whole lot of really good work that people will blame me in the comments for an hour anyway um but I guess my point here is that we're now entering other applications for graph structures which needs sort of a different alignment so uh what started like the schema August internet search for me where I'm coming at this is kind of human behavior so we talked about conversation but like you need something that sort of encapsulates kind of human behavior characteristics if you want to sort of interact with human beings and and know how what's appropriate what the guy grows are perhaps around because the phone that's fascinating me and I'm possibly Australian topic here is kind of why I wonder whether the graph will not only be uh key for the sort of language understanding but I'm sure there'll be graph structures for behavior and other things I think this is a great segue into something else you've posted Tony of uh the attention mechanism how similar it is to the graph yeah so some controversy over there so but but yeah let's let's maybe even um backtrack even a little bit further from that and let yeah let's talk about like text and data so at the moment we've got this idea of like I have structured data and that's what most of them are saying Enterprises have a lot of structured data so all these databases sort of all this data we've also got quite a lot of internal documents that are kicking around and then and and then out on the web there's the web of documents we've got all of this text and that's called unstructured Data but you you run a large language model over the top of that obstructed data and of course it's not unstructured there's a rich semantics of the natural language that is everything all the intentionality of the people and with a big enough uh neural net and deep enough neural network it turns out that actually you're able to learn a lot of that structure and you're able to get a bit of a kind of world model going on there as well and then on the other side the structured data as we were talking about before that doesn't need to be any longer trapped into these kind of boxes of this is the relational data now we've got this Rich structure of a graph and obviously a graph is just a type of network a neural network is also a network basically we're talking about quite suddenly we have these kind of like similar uh structures so then the question is well can we take the structured data and the unstructured data and and what can we do to bring these two things into harmony uh with each other and I think that's very much something that's being worked out at the moment but so let's perhaps talk through like some of the ways that that can be done and the first one would be like I'm just doing ground truths so like one of the things I've experimented with is you ask a question into chat GPT and but you say give me your answer back in a Json LD graph yeah it's seen it's seen so much uh Jason LD out there already it knows it very well right it does the entity recognition uh Microsoft here and you're talking about blah blah blah blah and then it gives you about that in adjacent now the graph and then if you kind of pin it down and say well okay look when you're identifying Microsoft I'd like you or any company or any entity that you recognize use their website there's the URI or use like one of the open source uh you can use their Wikipedia use the Wikipedia page about that entity straight away then what I can do is I can have my own information within my company and I can use that like a Rosetta Stone to link every entity I know about to the entities that are out there on the web so then you've kind of got a general you'll go so the way to see what the large language model is is a compression of the web and they basically taken all of the information on the web and they've compressed it yeah so then I can put a question in and get a graph back it can have well-known Uris to do the entities in and then I can reach into my own company's data and this is where the kind of Json LD comes in and having your own schema.org because if you've mirrored that structure so that's the structure that's working out on the web they've got a large language model they've got Json LD it's completely there's no limit to the amount of information that you can put in there and Link there if your own internal uh company structure it's got the same thing now a graph has come back from the general knowledge of the web I find all my kind of connection points in with any of my local data and I can now bring a whole bunch of data that I've done watching with anybody else and merge those two together now I've kind of got this working memory graph basically that's representing a code back to your point earlier um potentially getting your company data into that format you could even use a large language model potentially do that right so I was trying to ask what's the what's the two links because I'm going to specify a model that I found reason where they did an embedding of the graph when finding the model but it sounds like you're talking about two substitutions API in them or yeah so so this is where this is where there's various ways of bringing these products together so so we'll talk about those kind of like Ground Zero approach that you when you say to the large language model bringing back my data bring it back in the graph use these well-known Uris and children to uh and then I'm just going to use those as anchor points yeah and I'll Connect into my anybody can do that that's that's kind of like your that's your kind of Base one level then you've got Vector embedding databases so you could basically take your company information you could take a you can create a vector embedding like per subject for instance and then you could uh so I probably shouldn't backtrack or a belly so um when a large language model is compressed it's information about the web and when you hear about GPT being like you know parameter or whatever or whatever that is basically a long string of numbers and that long string of numbers you can think of it like a kind of coordinate in a way in an ie sort of way so if you imagine if there's just two numbers it would take you to any point within a graph and if you if you imagine it now like three numbers then that coordinate is going to take you up up and down as well and basically if you imagine every word now predict that's coming in the sentence and then I feed that word in now that word is going to get one of these embedding vectors and that's those have been around for a long time I could take the embedding Vector that is King and I can minus mail out of it and plus female into it and then I'm going to say well what embedded Vector is nearest to this one and it will take me to Queen so I can add these numbers this is the thing that took me the longest to get my head around is this sort of multi-dimensional relationship structure that's going on right that's what this Vector is essentially it's kind of a Einstein but anyway is this kind of linkage in multi-dimensional space of one thing to various other things and your point about compressing the internet is it's if you could think of if you could visualize the internet there's lots of different sort of points that is essentially what's then kind of describing it as if and so compressing it it's but it it's sort of mind-boggling sometimes if you look at some of these things for the first time it's well especially because like like open AI they are particularly quite a lot of their early papers about sentiment neurons say like you know uh they were just doing next word prediction and then they had a neuron but if you tweaked it around within there then it's going to be either like positive or negative uh sentiment based upon it so the the implication there is that some of these embeddings they're they're abstract conceptual World bodily uh neurons and and I think that something was still argument but to your point that's why platforms like um elasticsearch and Pinecone are taken off again because these vegetables are proving so valuable to solve hallucination language yes like I said you can take your databases embed it every day when you're update and then they've got a truth that sort of truth but they don't want to stop yeah you need to have to face up against another Jones model but is going to be operable we're interoperable with it right so that's why I really like obtaining the saying around Jason Aldean I guess maybe it's something controversial when you said it but now like looking back we've got kind of it feels like that's you know not I mean it's a it's a build it's about something to make a time but it kind of feels now you've explained it like that but that's just very sense a very sensible way yeah so well so imagine a date so some of those conceptual understandings one would be able to map back to schema.org right so where schema.org has got a concept that's modeled in schema.org it's got a whole bunch of data examples out of there it's got all the text that's surrounded about that sort of thing so now if I'm in my own organization and I've based on my model of schema.org model and then I've got all of my own data items in Json LD that are pointing back to there and I can make connection considered by getting the embedding vectors for those sort of things it's like we're speaking the same language now the main value being you can Leverage public information so whether it's in finances and a fine example like what's inflation that'll be on the skipper.org on the web and you can just pull that rather than store yourself well it's more kind of knowing what the world is like say you know you can just oh you can you can say you can just put in question you can you can put in questions that you would ask to a human and that kind of Common Sense understanding of how the world Works based upon the internet being put together which is then mirrored with your hardcore data grounded into your hardcore database you can you can do checks on your data as well yeah right yeah well so the the kind of vision I put forward to is every organization behind its firewall every all data everywhere accessible via this semantic layer so you create this semantic layer which is based upon schema.org but it's got all the specific terms of your business and then you say to each application you will publish sorry I know that you don't want to but tough you will publish your data in Json LD representing the data that you hold within there this feels very much like the Jeff Bezos memo The Immortal one right that says Thou shalt service it kind of feels like maybe I think it is a bit like that but this is this is this I believe is the crunch moment yeah in the exponential curve the conventional Western that you can't have one of these shared models it's simply too hard and most of the banks have had a go at it at one point and they've failed but the difference now is I think it's a good Vision behind it failing fail it you can't fail at this now the luxury point of not having one definition of what your customer is or a base model of what a trade is or a base model of a track is if you can't get your stuff sorted out you're going to be in serious seriously that's not just the channel engineering suddenly that's getting a lot more interest because everyone's wanting all these nice features and use cases garbage out so you need to get your data and you see all these movements around data mesh data contracts and they're all they're all very good but actually guys check out what's going on on the web there's a model out there it's working don't bother Reinventing the wheels to take that and use it and then extend it where there are specifics that you need to if your industry and for your particular organization but try and align as much as possible to be outside well so it's interoperable makes make sense yeah um so that's kind of where we're at now Tony I'm really interested to sort of get your take on where you see things headed or are there things that are really capturing but yeah we'll rephrase that slightly let's start by sort of looking at things that right now are capturing attention and I'll ask you a little bit about where you see things going in the future so what sort of caught your interest recently in this space so what I'm super excited about at the moment is like the rise of these open source large language models so I was playing with them on the weekend yeah and they're getting pretty good now aren't they so like you've got like uh the the uh the red pill armor and stuff pajama and stuff like that that now we've got these open data sets so what I would be really Keen to see is where can we take one of those open source large language models can we take these principles I'm talking about here and build that in from the start like could we take Wiki data partly train uh create a data set in red pajama of like an output from Wiki day to train it on Wiki day to um you know so that it it would now have a we would now have a Knowledge Graph embedded within a large language model so some of the so anything that we can do to like sort of squeeze uh the difference between these two worlds of like the text and the uh and the factual grounding you know can we take some of this can we take schema.org and do more of the um can we do more to train it there so I think that's quite interesting and then like this whole idea of tologies and can we take something that's going on within the long large language model and then relate that back to the ontology at the conceptual level can we zoom out from the conversation a bit like kind of what you're doing by the sounds of things by creating your ontology about the conversation it's sort of able to sort of do more structure as I say so I think it is I think it is I think it is more structure and it's a bit like um I think the human memory works this way so if if you think about like like the uh Butcher on the bus so you you see it you see a guy on the bus and one part of your memory is like recognize his face from somewhere you're on the kind of fuzzy part so this is like the large language model part now it's trade on fuzzy textual data I recognize and do it is he a teacher at my kids school is he and then the second part of your memory kicks in and goes no he's the butcher that's where I recognize him he's from so I think it's almost like what we did was give another abstract layer of function or goal so the weakness or perhaps what we've identified with graphs is yeah it's good to put in the sort of structure and makes it easy to query and there's like some models are better like Recreation models but it kind of exists but we're looking at is more um to that point like what can we do with that how can you actually manipulate the graphs to make it more um useful yeah that was more of a random thought I think you just need two points to click together touch your example yeah so that's so that that's how human memory works and I think maybe we can we can start to move to something similar with these things the other way of thinking about it is like it's like fast and slow things so again it's like they've got the large language model it's got this World Knowledge it's trained on text and then we've got the kind of ground truth facts yeah which which you need which is exactly the way I was saying I think we need to integrate this this thing is a component of the broader system we're trying to build and that that sort of thinking slow piece is the kind of organizational history and data and other things and integration architecture with perhaps the latest feed of Market data or customer information Customer Events whatever those sort of things need to be sewn alongside the bit that's processing the you know the language component I'm sure it will change when these models like say in the future when these models exist to have other features wrapped around them I'm sure will be even more powerful but for me at the moment it feels like you need to sort of cook up the right blend of pieces to make this work well yeah and quite a few people seem to contact me who are working in the biomedical smoke okay there's a big Ethel at the moment to kind of take over the science papers and then you know somehow with a large language model get increase the knowledge that's there on this on this kind of scientific stuff so my message for those guys is it's just always at the moment where you've got already a rich set of ontologies which are human engineered kind of ontologies that have attempted to create a logically consistent understanding of that can you use those autologies in concert with your large language models to to kind of again kind of ground them out so it also give us a good example like what we were looking into so one of the one of the weaknesses of language models especially in the conversation AI space is um stochastic parrots so point being you know if you just say credit card or a word how we humans think is what you're talking about Chris what do you mean credit card well what's an animal just do a portability and just explain what it is yeah yeah um yeah that's what we figured grass is really good for us again if you're if you're missing context that's how we human brains workers go okay well what the news that I'm missing I need to feel those in and that function then you can feel that Backpage there um seemed to work yeah I think that's it it's your point it's like that it's not just not just information extraction but you can actually put the functions in the graph to make the language model um but more effectively yeah yeah and and now imagine an organization that has basically established in semantic layer has got all of its applications publishing there the key parts of the data they hold into that semantic layer has maybe trained its own large language model obviously a great expense but taking one of the base models done a customization overall we see that but the open source demonstration is coming down yeah it's kind of fast yeah yeah yeah no it is so you take all your documents to map the documents into the kind of semantic layer as well so now you're basically training on all of your documents you train it on uh your ground into your actual data as well um and then it doesn't seem inconceivable but you you've got a window where you say oh I want to do xyzed functions like autoship act as a task manager but that's what I've been into recently yeah there you go for you you've got an assistant that knows your bank but the planet ground one is is actually something really quite simple and glamorous and yes you'll be able to use AI to help you do this process but actually it's just to get your data sorted out get you can download schema.org from it's just a download download schema.org and then start the painful process of trying to get all of your you know really it's going to need anybody watching this now you want to get some some way of getting to your chief executive and delivering them this message which is that you know we need to be getting our data estate sorted out now so we need to we need to get this downloaded and we need to begin this painful process it's political yes you know it absolutely is and my thoughts are going to kind of article data models right which were a thing yeah it's possible very difficult politically but this kind of sounds like from my time so canonical data models where um an attempt to try and get organizations to have one sort of standard model that you would probably map business unit data into in reality but so for example when you we have an integration bus or Enterprise service bus type thing in the organization you want one sort of standard set of terms or data sets going across that bus so you can sort of publish And subscribe so it's kind of required if you want to do event driven architecture well because in order to publish events and subscribe to events you kind of need to have like a greed business language the trouble with it was is that you get each business unit going well now my definition of trade or my definition of customer is the right one so where I'm kind of kind of coming to is this does feel like it's going to be quite painful exercise however the point we were touching earlier is there are no there are new tools that our disposable now which is we can perhaps automate a lot more of that translation um and perhaps you can have um data sets that aren't aligned as long as you've agreed on what the standard terminology took schema org is you could have something that perhaps on the Fly is doing some of that translation as long as you've agreed the mapping so I think the problem we had with canonical data models was it was just almost impossible task and it felt very technical like what's the purpose of doing what's the business outcome it was sometimes that was quite difficult to sort of see I think technologists you know enlightened Architects and Technology strategisters could see where it could take the organization the trouble is that perhaps wasn't so always hard so yeah whereas now I think it's Tony's point you kind of see this the the potential of of kind of the graph sort of structure and what it can do for you this perhaps yeah the lights will perhaps go on to go right I can see why we need to invest in our data stewardship governance well the pitch simply now you can make chat YouTube talk about your data yeah or you can do tasks yeah yeah that is essentially yeah and and really also uh do or die and I so I I hope I hope that people do revisit that because in the the back of my head is this worry that basically what happens is a lot of people don't don't sort it out and then we just get this kind of like uh Wipeout of lose a load of diversity within the kind of economic system uh which will because the tech sector and tech savvy organizations that get this right yeah a few early movers in every industry it could be like you know like One Bank could do this one Whatever could do this they'd do it first the other day because of the human aspect to it because of the changing culture that's going to require to do it it's not yeah okay the AI is getting faster from Fast we can chuckle the AI at it that you want if you can't get people to kind of agree and collaborate and get the structures going then um it's not going to happen and for those that it doesn't happen for I just can't see how they're going to see yeah I agree it's very similar to the whole Singularity item of the problem with this AI technology is before large organizations could feel comfortable not being first moved that's not really a case balancing energy of your first you win there is like you're not just a second behind your your years behind at that point that's only going to perpetuate histology again we're entering as you say the the six mental curve of capability um yeah I completely guess that's my concern too is um if you can imagine like if you could be able to find your coach or if everyone has their own personal financial advisor I'd sign up to that bank in a heartbeat and all of a sudden like three and a half year behind you then I'm going to switch back no that's just one example yeah he reverse you can speak so this has been a fantastic conversations I specific words the things I'm sort of taking away but I'll be interested in some of you all sort of final uh thoughts a pair of you as well is that this is a genuinely uh game changing sort of moment and it's not just the Overlook the computer can Now understand and talk you know talk our language there's actually now going to be a whole bunch of other things that can be unlocked behind behind this as well so it's got this sort of one-trick pony that you know is overhyped and and then we sort of forget about and then the second thing is we go back to the importance of uh you know good architecture essentially which is good data but also good change management around getting people on board with the journey the mission to sort of become more data you know driven organization and really be good uh data you know citizens when it comes to sort of you know managing our data stewardship and all that good stuff so those are the things that I'm sort of taking away but Tony I mean where do you sort of where do you sort of see this earlier what's your sort of thoughts on where this is headed um yeah so I see lots of I see lots of change uh going on I see rapidly evolving you know the the tectonic shaped uh technical tectonic plates are going to be shifting around in a in a in a increasingly fast spin um and I think that through all of that kind of fancy stuff I think the message should be to all organizations is just concentrate on the simple stuff get your data get your data organized and get it in shape don't get don't get too distracted don't get too distracted by all of the wears and everything that's going on you've got one simple Mission now yes use yes use the AI and pour it into that but you've got to get the message to your chief executive the chief executive is going to realize that this is probably the most important thing for any organization right now since the internet gave a lot of wraps yes I mean Bill Gates the center Bill Gates um where you you know you were involved in the making of the PC and the internet where you see this moment with GTP is it where does it sit on that and he's like big as big as maybe bigger than those so as big maybe bigger than the internet or the invention of the PC and that's someone who would this is definitely this is definitely real and there's going to be a lot of movement going on in that but that but in a way to stick your blinkers down get your do something that seems kind of quite basic and boring and and it's certainly going to be painful and politically awkward which is just get your get your semantic layer set up so get your data and your documents nicely organized and then when all of this kind of plays through your position you're in the go this is this is the ground crown jewels that you're sitting on no one can take that away from you it's your information know what's private by the way that's also very important yeah let stuff leaking out earlier that shouldn't be leaking out yeah because that you've got that you've got the experts you've got the data that you've built up lock that down get it organized get it consolidated yeah Chris I feel like I should I don't think just to keep the message pure honestly um find the right people to do it yeah don't talk yeah my group of strategy and Implement that is also incredibly difficult to keep it simple too so we don't have enough people who know about stuff and there's some Wizardry around it which is unnecessary it's not actually as complicated as it is well that's why I was really Keen to do this sort of mini series really is to kind of make sure the architect Community understand quite how game changing this isn't the the thing is sometimes folks in the Enterprise architecture space and roles can be a little bit like oh yeah I've seen I've seen this thing I've seen you know oh yeah yeah there's a there's a bit of a oh yeah yeah it's the next type thing I think sort of pricking people out of that bubble say no take this one a bit more seriously it's not it's not blockchain it's you know it's not anything it's not web free it's not ironically and then yeah it is web free but the real version so I I hate the fact that the web free has been hijacked by by people that are kind of trying to take it down a different part for me this is what went three is it's it's it's the internet that's uh compatible with machines as much as it's compatible with human beings and with that uh I'll just thank the pair of you very much Shane Charles could make it I'll perhaps get Charles thoughts on this topic separately um but yeah and if you aren't already um subscribed uh do check out the architect Tomorrow YouTube channel we're also on embarrassing audio podcast services and Tony here and Chris also have published some great blogs on LinkedIn so do go check those out if you haven't already and with that we'll see you on the next one thanks guys foreign

Channel: Architect Tomorrow

Views: 5,460

Keywords:

Id: lhRYnZS7yu4

Length: 47min 34sec (2854 seconds)

Published: Thu Sep 14 2023