Latest / Skit om AI / 3: Vi pratar databaser i en tid av AI med Jens Nordström

Transcript

3: Vi pratar databaser i en tid av AI med Jens Nordström · Skit om AI

• Sebastian: Hej, jag hoppar in bara för att säga att från och med nu de sista fem minuterna eller vad kan vara så av någon anledning laddades inte Jakobs enskilda del i det här upp. Så ni kommer ju då höra att jag kanske frågar Jakob vad tänker du om det här och sen får vi inte mer kontext. Så det kan bli lite off-kontext de sista minuterna här. Men försök hänga med ändå. Bra, tack för att ni lyssnade. So, hey, poor Ed, hey, Jacob. Jacob Kroon: Hey! You're most there. Sebastian: ⁓ And... But it feels good, it feels good. How do you feel, Jakob? Jacob Kroon: It's rolling on. It's really good. lot is Sebastian: Great. Exciting. Today we by our dear Jens on my company, Skjärna Fyrkant. I remember in the previous episode we talked a little bit about this with security and databases in AI projects. was for Jackie, you were building your own CRM system for your business. Jacob Kroon: Yes, exactly. ⁓ And that was like new information for me. I've read a lot about it, but you can still say that you're not an expert. What I read was that it will be very important for big organizations with HR and more legal... I have not read that much, so it would have great to learn more about Sebastian: We'll have to into that together with and what Jens has found about us. But Jens will have tell us a bit about himself. So I say too much. Jens, how are you? Jens: ⁓ well, it's just nice. I'm really excited to talk to AI and bring down others in the deepest rabbit hole I've ever put my foot in. So that's to be exciting. Jacob Kroon: Yule. Sebastian: Nice. Rabbit hole, what would you say? Jens: You are enthusiastic, we all are. You like to learn new things and learn about new techniques and everything that means. But this is probably the subject that I all the years. This is probably the thing that I was like, wow, it was really deep. And it was very complex, very fast. So that's where I stand right now. Then you you have been a lot on AI now and self-taught yourself for a few years. So this also like a natural step to least get a good grip on how it works. and how to work against it. What I've when I've worked with AI is that more I understand how the LLM is built to respond to me, or the tools that exist to work with AI, to help it on the the better ⁓ context and response I've gotten back from AI when I work with it. So then I in a third step that If you this part too, think you can get really stuck in your general building and mess, if you understand the whole chain. Jacob Kroon: If I may interrupt, you have learned how respective LLM is trained, like shutgpt respond this way, clod respond this way, and which one use now has its own way of responding based on your prompt, but you also to prompt in a special way so that security around what you are should be correctly implemented. Jens: Absolutely. That's how is. You make lot of own assumptions, I when I look at myself and I'm building with AI and I'm to code something or whatever. I lot of assumptions, things that I think are self-evident, I get on myself that just... I've never asked for this thing, or I just assumed it was. Of course you have to put a safe word here. Why wouldn't You've had to few back over years and say, okay, I really to make And that's first thing I do with you, the safety aspect. How should it think and act in the safety relationship? Jacob Kroon: But is that the first prompt you do that no matter what I'm going to write here next, you to leave these points? Or how? Jens: Nah, I usually do like a... almost like saying that you should build a skill for an agent AI. You build a document that is alive and that is developed with the portal or product or the idea you have and that should always be to when... You can also prompt when to the document. We're going to build this. Always read through this document first to ensure that it is secure. Like doing the same thing with a design system. Always the design system when developing a new page. Always check the security document when building a new connection to something. Sebastian: That was a good comparison. If you to design something, always want to an in first before you look at what type of brand book you want to what UI you want work with, so that it every component. Then you have the same type of guidelines, but safety issues in infrastructures. Jens: Yeah, but... kind That's pretty new to me as That's something I'm trying out now. But I think it's worked out well. And it's a come from design system. We should have a specific place where we can always look at each Kind of with databases. It's the same thing. Here we have a node for this, here we a node for this, here group this type of data. So works a little little bit. Jacob Kroon: Would you that is exactly as critical to work with security in any project you work in? Jens: Yes, basic principles. But then it on how far you it, depending on what computer you're working with. Sebastian: But Jens, can you give us a little background story about who you are, what you have done and why you are doing what you are now? Jens: Absolutely. I work with Sebastian at Framafrikant. I am the chief consultant here and before this developer. I worked in a front-end for few years before I started at Framafrikant. And what happened in my development journey, I got chance to be the here, I it's really fun. But I have been to development again now. AI came and helped everyone interested. And what AI did with my journey was that as front-end, you had the whole concept. You knew how the backend worked and what you needed to for in the backend to get your front-end to work. What AI did for me was that now I the tools to build that directly. I didn't need another team that did it, AI could be my backend team. As long as I the directions and I create the thing myself. And made me very interested and thought it was fun to start building things again. And Sebastian and I, technical nerds, started to lot of and took it on the road. And the we up with a of in that we... We want to build our own stuff. that's my research now, you call it in databases ⁓ and as a whole. So goal for us is to build a master database. Here is the company's ultimate secure brain. Which we can then connected to end user portals or management systems or economy reports or like... If you with that brain, the is endless. Again, front end or what you interact with as user, that's the easy part, you've realized now. Building this box where everything exists. That's where you're put the time, that's where you're going the energy. That's also the most challenging part. Jacob Kroon: Fun! How's it Jens: So, so, but then, I'm a... I have to say, it's well. It's insanely complex. ⁓ I have some talking points to describe the different layers and how deep the rabbit hole goes. But I have tried a lot of different... You have company data. But what company data? It could be PDFs, sound files, pictures and movies. You don't have any structure. And other things can in the future as How do you for that? And I'm like, in everything I do. Okay, in ten years. How will You really don't to get stuck in something that makes you... Jacob Kroon: So, mm. You mentioned something very interesting there, but when you build yourself in a corner, is that even possible in today's situation? non-thing Jens: So here, I have experienced and discovered that even though you as a developer or a driver think that your boat is well and is going on the course, you were lower the water surface a little bit, might be lot of holes in the scut and there are a lot of sea bubbles that you have to screw away. Even though you're and changing, AIN does that just as far as If you're not really reworking and reworking, it takes a long time to rebuild something that you've already implemented. You can get it goes very fast, many cases, depending on how complex it is. But I've experienced the whole project itself. Jacob Kroon: Yeah. Jens: not up and converted as fast as you think. There will be a lot of legacy stuff left. But answer your question. No, can build from what you have. But it's mostly about getting a clean product. It's also security thing. If you lot of legacy code and legacy stuff that you don't know about, you also a security risk. There are many elements. sit on when comes to that stuff. Jacob Kroon: It arouses many fun thoughts. Exciting scenario. Yes, It will haunt me for time. How to handle it. No, but I'm in lot of projects of different kinds. And almost, I would say 99.5 % of all these projects, you start with one demand and end with another. Jens: Mmm... Marrømmer, you Sebastian: Don't wanna Jacob Kroon: What you want will always develop over time when see things start happening, see a product in of you or you see a new web that you can click around in and then you think, oh shit, I didn't of that, I didn't of that, I didn't think of that, but now I it. And then you can already have built yourself into a corner So I would say that it gets put in things as soon as have launched something. Jens: Mm, absolutely. Jacob Kroon: Taking things that you thought were important in the is not as important as it launch. And how should you think about then? you rebuild from scratch, because it goes so fast and take the basic ground or basic skills and just make a cold turkey and then you can rebuild the every time something should Or how should you think? Jens: That's how and that's how has always been in traditional development as We don't sit with development as company in that way, but we also projects and what the customer thinks they want or what we think we want, depending on the project, the goal is over time, you find new things and so on. So that's something you... But, yeah. think the bigger, the faster and more AI models will I think we will get to a point where we won't have build the system ourselves at all. It will be AI that is system, no matter Right now we're just gliding on... ...frontend stuff to get what we want. But meantime, I don't we... It won't be at all. you to if you work with safety, but it's the I don't know. more for. Jacob Kroon: Yeah, there's probably no answer to that one. ⁓ It feels like lot in my you Sebastian: Yes, but it does. You're thinking about this as much as you As Jens mentioned, for our company now at Stjärne Fykan, I'm not... I'm not... I'm absolutely most excited about having a master database that you can connect a lot of smart functions and to. That makes it that we can build small micro-flowers for exactly... I talked about it earlier, think our CRM, AppSales, is a good CRM, but it's about the surplus for what we need and too expensive. We can build something better, but I want it to talk really simply with all the other systems or UINs that people are interacting with, whether it's a customer or another colleague on another branch. So think this is the But I have respect for the it's not an easy question to solve. And I haven't cried down that much. But we'll get into more details about how Jens thinks that we should do this. Jens: That's interesting. I think the is to build the underlying flows. They will be true for us as long as make structural changes in company. Then what limit we present is the same. But I think the time we should invest is to build the database and its function as well as possible. And then these micro flows that ... to meet specific small needs where we save time and get out a lot of Then, if you're going build a portal for customers or employees or whatever they can work with, that can certainly change lot over time. the database and the are more consistent in relation to rebuilding the product that we were talking about. Jacob Kroon: If you were to build a CRM, it be like a layer that talks to the database and presents it you want it And then the flow between them. Sebastian: Maybe need to explain ourselves, Jens: Absolutely. let's look at... I don't how much you know how this monster of technical terms and concepts was when it came to an AI database. what it actually means... Have you read anything about that? Yes. ⁓ Sebastian: Alan Tupsel. Yeah? Jacob Kroon: Enlighten us. Sebastian: Yes, Jens, this will be be as a video as So if you to the and show us something, and also those who choose to and listen to this, if there is anyone out then go ahead. Jens: I don't any graphic material. But we have to rewind to the days when databases were How have we lived in databases until AI came? And that's something called a binary database. It's one of kind. have no... It's very... specifikt. Vi skriver jag vill ha årsrapport 2026 på vår dator. Vi tar de det ordet och söker specifikt efter det på vår dator och får tillbaka ett one-to-one svar i princip. AI jobbar med någonting som kallas vektor. En vektor är det ordet, eller vi kan ta en enklare relation, vi kan ta hund. Hund har i en stapel, tänk ett stapeldiagram. very much fur. It has four legs and it tame. Also, a vector these things differently, depending on how much them are true, you could say. And then we take account a wolf. It is also fur, a lot of fur. It has four legs, but it is not tame. Which means that in a vector database, these are very close to each other in relation. Not because they look same visually or because their words are completely different for each dog. It doesn't at same, but because they are classified in a vector rating, they are placed close each in a vector database. And a vector database can be seen as a circle. And every star is an object and depending on how this rating is set, are placed close to each other in the Are we ready? Sebastian: Hmm. Hmm. Jens: Nice! And this is fantastic! We can save a lot of it's really good scoring, but the of a vector database is that there is no hierarchy at all. And that becomes very problematic when we are working with corporate data. Because then we want to know who is in charge of who, who is right in what. Yeah, okay, we can find... Sebastian: Mm-hmm. Jens: Here is a meeting of that happened, but all the and protocols for that, because we don't this one-to-one search with the we want the meeting notes. There is no relation because they have been in the vector scaling differently. So that's you to implement something called graph search as well. And that's where the of is now, that you have this big space to collect a lot of data. But then you also have graph search. And graph search is the opposite. It's based on nodes that manually draw line between these two nodes, and this meeting, and so on. So they, in combination with each other, create a... a that both people with words and with a datalog can search for. And did you know, just a quick anecdote about that, that I've now, I was like, what? An LLM, Large Language Model, is called, can't interpret text. It rebuilds words to zero. Jacob Kroon: Okay. Sebastian: Alright, okay. Jens: which also later becomes vector graphs after zeros. This is so counter-safely that it doesn't It's a large language model. But all the interpretation it does is in numbers. How sick is that? Jacob Kroon: So all output is actually ones and zeros that convert to letters. Jens: in principle. Sebastian: You will never understand that. Jacob Kroon: Bye. Jens: I was struggling to google it, was can this actually be true? Because they're out there in the Large Language Model, it sounds... But they work in semantic meaning, that's it's Complex stuff. Sebastian: Yes, you can say that. And this with graph search, vector database together with graph model, convert the into a graph, I don't really Jens: There is a certain amount of in this, we didn't mention but it's a vector. The search is built with something called RAG, Retrieval Augmented Generation, it's the current standard. Sebastian: Hmm. Mm. Mm. Jens: It helps AI to navigate in this space and find the one. It's a graph, that's we've decided. Where we've said that product A is connected to product B and product C. We've that manually for really critical data that must have that traceability. Sebastian: Hmm ⁓ Well, you'd visualize a database that, if you were see nodes that have stretched between each other. You understand me correctly. Yeah, but you those... Jens: ⁓ deep soul. Awesome. No, but then you a dream with a lot of stars and stuff. So it matches these things. look at Jens is someone's boss and then we all his... one-on-one conversation with the employee. finished mapped because there is metadata, is... And metadata is, for example, something like, time, when it date, with whom, which team it to, who has access to it, and things like that. You almost set up graph data too, so that you can do this mapping on a site. Then we come to metadata itself, it's like one of the data points in an AI database. So these small tags, like what, who created it, it what map is in what time do have to do it, etc. Super important when building a RAG AI-based database. Jacob Kroon: When you say manual in this aspect, what does that Are pre-determined rules that semi-automated? That someone sets the Or is there someone who sits there and pulls the between the Jens: As I understand it, it's 100 %... I haven't done it myself, but as I understand it, it's 100 % manual relationship. Then you can probably build scripts and like, seeding things that say that, as long as all the things that are between Jens and Sebastian, they should have this connection. Or if they are in this map and... The Sebastian and the editing possibilities of the It's possible to a framework around how works. every... You understand it yourself, a company will a lot of It will very... A of manual work, Because that's the thing, ⁓ to take the data source, curate and prepare and adapt the data so that AI can be used later. Only that job is your big job. You can't just snap your fingers and say, now AI can search here. That won't be good. Sebastian: No, and I can give concrete example. We have our SharePoint, for your reference Jakob, I lot of the data that is our company is on Microsoft SharePoint, no wonder. All our files, Word documents, PDFs, Excel, whatever. And there we have built up a structure from ISO standard, quality and management work. Top, great, it's literally organized and so on. That's what you lose grip on over time. It's just a feeling. No company can keep the structure that the wants in such a system. Ain't gonna happen. So that's what we're trying to over time. And then there's double messages at different places. That's the next problem. Say that we can do this with our supplier. Supplier X. These the people, this what we buy there, this is how the routines Perfect. Next time we're to update that. Now it's the new routine for the supplier Rex. We've changed the person, we're going to this this way. Maybe someone makes a new document with that information, loads it up in the same place or somewhere else. And then just now, now we're going to have an AI agent here on our internet who's for info. That's awesome. And then we the co-pilot. And then you ask the co-pilot and then you an answer like... You have no idea if this old or new info. It starts referring to documents, but you still have have the knowledge yourself and understand that this is old. And that's where the metatags are again, so if you could classify it well, this is, I still have this info, but it's old info. This is new information. So it understands where it should go. Jacob Kroon: It feels like SharePoint should already meta tags implemented based on when you a document or information of some kind. Jens: I'm not I don't think it's very extensive. But I it's creative and who owns the and such absolutely correct. Jacob Kroon: ⁓ Sebastian: Yes, they But it's still... Has anyone been in the lately? Although it's old, it it's changed a certain time. So it doesn't work great yet. Then can make more settings around it, but there it's a bit overrated and I can feel Jens: We're jumping into pre-development but what you're describing is related to the type of or AI database, the fear, or Silent Breaker. Embedding drift. Embedding drift is when a database, we've built a database, and we've said that it's Claude's 4.6. who will handle retrieval and work with that database. It has its preferences, its trained data. Because you should also that without RAG, that's a good little thing, without RAG, an AI always use its trained data as a memory or as a fact. RAG forbids trained data to be used and taken away or filled in with guessings, like what we call halidization today. If there RAG, a, if you have built it the right way, course, then a AI to use a structured way to collect data and present it to you as a But Embedding Rift We have said that CloudSanit 4.6 is the one that will pick up work with this data for us in our KB or whatever we have built for the product. And everything since we update that data, we may change name or change branding or change staff or our data evolves at quite large key points. the data will become less less accurate over time. And if we also change and update the model we work with against this database, it will also become less less accurate, because we have built and structured our system based on how a specific AI looked at a specific time. Therefore, is recommended in industry to never change models when you have set it well. And if you to change a model, always to do a rebase. You have to read the whole database one more Read it on the net. Map it. Or whatever you want to Sebastian: ⁓ my god. Jacob Kroon: There was so much information there, so have to go to the I'll back. Jens: Okay. Sebastian: Hahaha, okay. But we can continue with that thought. That's really scary, Jens. When did you this idea? Jens: You Mm, I know. like a week ago. But here we are on... This is a lot of research down here. And this is a huge problem today for everyone who works with orchestras and this is also about... It doesn't in a week, month or maybe year, but... Sebastian: Åh herregud! Jens: If you continuing to in and not find a way to either clean or like rebase or rebuild sections of your database over time, then it will be problematic. it's a bit like with all this, you say, damn, we build a lot of stuff that you don't follow up on or see. It's maybe like, how should we structure and maintain it? support what we have built and make sure that it delivers constantly over time. Because data is alive. Sebastian: Yes, yes, yes, and that's we come in on the basic principles before we even into some kind of development project. We don't need to it AI, you develop with AI, then AI doesn't be in the But any development project, which de facto value does it Can we measure it in some way? Can we concretize what is for value? What the usability of this? Will people want to use this? Do even that or will we want to continue with our old stuff? What are the risks? I the security risks when you have a CRM or a customer portal, customer sensitive data. How do control that data so that it lead to loss value or leak? That the wrong person will get the wrong things. And then the fourth is what you're talking about now, Jens, this ownership. Who is going own this product? How is it to If you don't an for them, then it's... It's high risk to jump into a development project and believe that time is worth something. Hmm? Jacob Kroon: How well do you think companies today... How good do you companies today that work with AI are to handle security? Jens: There's one those. Yes, Jacob Kroon: Helt värdelös. Jens: I was in an education program a few months ago. then they broken down into different workshop rooms. Different companies all Sweden. And you there and talk and people show what they're and what they're I work at an MSP. Safety for us is very... close to Hans, it's everyday all the time when you talk about it evenly. And then you out there on logistics, where have logistics companies talking or... or... warehouse companies working with them so that they have an inventory system and stuff. There you don't that at all. It's not mentioned at all that it's a feature or a part of the product, but you just want get a function. I think you realize how connected you are to when you... Just having an internal product that you just use on your local machines is still huge security gap if you haven't thought about Sebastian: Thanks, Should we continue with the database explanation? We've through graphs, vectors, metatags, more the will when you build up a database this? ⁓ Jens: Hmm? Yeah, amen. You can say it we have... ⁓ We have RAG, and it helps AI to search for star more easily. But it lacks relationship. So it helps to create a graph part in the database. Then can also help RAG with metadata and tags like that. build relationships and better. the metadata, I say it again, is very important. Then, when get into the that when we take data that exists today and put it an A.I. adapted database, then need to do something called shunking. Because as I said, all text becomes zeroes zeros and then it coordinates or like a rag score and then it out. and then it will be to with the of the coordinates that it gets, we can call it, out in space. And when you do shunking, you take a PDF, then you that document, per piece or per section or according each meaning, depending little on it. are different ways to do shunking, but it makes the coordinates more precise. I'm looking for this part of the document, a small section that has a specific coordinate in this room. If I a whole document and say, it's over then my search precision worse. So that's shank up ⁓ text. based on different types of models, speak. There's recursive chunking, schematic chunking, document structured weight chunking, lots of different The that's becoming more more of a brand standard is, because our AI models have started to so advanced, is you read the entire document and let the AI analyze it, and then chunk it based on relation. Because could be like you have a... set a piece about a man and then in piece number two you talk about that and then you don't know that it was a man you talked about in the because we were in the piece before. But for AI to analyze the document, I that I can't stop here because then the text loses meaning in like shank 2. So that has become the new times and you it was not cost effective or AI was not good enough to do this, but now Jacob Kroon: Mm. Jens: can you now read in and get it chunked off and in before it placed in the you do like reverse on what you did from the beginning. Jacob Kroon: I think that's I've been implementing in my life in think it feels like mansplaining deluxe. When I write several messages in a of writing that one, I write... I have some kind of name for a previous message that I've sent. Like when we did this function, I the function to what it actually was. Or that I have like, dead... a prompt earlier or have a key word in it, I always that key word instead of what we did earlier. It feels like you're brutally transparent in everything you do, or you write or say in the comments. It feels like that... Jens: I all a man Jacob Kroon: Because I feel that context is the key to everything you want to You work result-wise in my projects. Jens: Absolutely, really that. And then I get so terrified when I talk about earlier when I have change and get things Then I've been so damn obvious and checked off all boxes and then I'm like, you do a security audit on all this and find a way to a security master and identify all the security holes that this product can have before it goes out on the Then you get list of 50 points and you're like, how? Jacob Kroon: ⁓ Jens: I've through this. How can you find this anyway? You can do ten of these after each and just spot the problem and go, oh well, now we to do all this It's so cool that the little reverse psychology to make it and try correct itself. Sebastian: Yes, because you should always leave the and that's what I've from customers who working on some development projects. Claude, or whatever model he is, is so extremely inclined to solve what you want in the simplest way possible. So that's why it's like he can always find a in himself, because he's just trying to... She picked up one plank from the screw here, and then she has spiked again with that plank here. And now I fixed it! And you feel, yes, damn nice, but how does it out there instead? You didn't the plank there, did think? Well... But Jens, we continue on this, if we take it very basic, take Jakob as an example who is a CRM system for his company. How should he think about He will have a database in the and then there a CRM and there will be different users on the company. Now this is just an internal system. There no customer who should be logged in there, but we may still want to certain rights issues in this database depending on who is logged here. Jens: I'm on the Sebastian: How does that Jens: First of all, you to... Well, can rebuild anything, but it's important to go into a project like How should the structure look? Is there Or is this system for sellers where everyone should access everything? Then it's super easy. You don't have put so much weight it. But do you want to able to segment information? In the system, you have to you from the beginning. I think that's one thing. Then you also to look at, okay, when we build this system, should we take in external libraries, or is this only based on internal data, like business data that you manually have entered sometime, or should you take in, there are probably APIs for all companies, example, such things. Then to at security aspects. Then it will mostly be text, guess. There will not be much image and video in such a system either. Jacob Kroon: Come on. Jens: Yeah, that's something you need to look at. Should we build in a way that we can have the documents and the documents searchable? And then, like the I talked about, when we ran for a you want to structure the database. The you can call it that. There are three big parts of There is the where everything is saved, the database itself. Then there is the free or the conversion storage that interprets code down to the database, which is called ORM. And then is the brain that tells what this storage. or send. Which is top-class. And there are two big ones, Langtjain or Lama. Or L-M-T-O-N-E-T-E-R or Lama. like the Langchain is more for agentic AI, which builds processes around AI, is supposed do functions, collect things or build anatomical things. And Lama is more like pure database retrieval. There is about 40 % difference between performance on them. So if you're just to collect data, you're 40 % faster with Lama. But if you want to be to build the agentic flow in your app, the bottom of the line. Sebastian: where you get and write data, like, you think so. The agentic flow should be in that database, that's ehm, yeah, but... It shows, you might have an API against all companies, it changes in management, or you have an API that... going go to the website and okay, this person has changed the company and is to get that information from an external source, into our database, update it, and rewrite it. That's called Gentic workflow, right? Then you want to have a long chain. Jens: Mm. Mm. Exactly. Correct. And maybe send an email to the who is the the package. And then you want to Langchain. disadvantage or the is that it's a predictable thing. It's a chain. A chain of events that happens every time. It makes it a little slower. When you're in Yemeni something, use thinking mode. It's Langchain. If you fast, it's llama. Sebastian: Mmm. Eish. Jens: Now it's not really like that, but just to explain the experience before the end of the sentence. You get resonating oi versus a retrieval. Sebastian: But Jack, now to the CRM, where do you the database now? Where do you building it? Is it super based? Jacob Kroon: It's super bass that we use... It's Loverbolt that we've built this series in. And they have an integrated super bass that has its own user that we only get access to on the left. Because that's how with all the that Loverbolt does. They have lot of connected APIs to their system that just... You can just write to it, I would like to in my app and say, but we have already implemented it. You just to click ok and then you to add any own keys or something like that. the super bass. Jens: But Hane, I don't know, I haven't built my Super Bass on the way we do Hane Access into Super Bass and can build and change things in there. Jacob Kroon: ⁓ We haven't been Superbass. Exactly like that. ⁓ Jens: That would be a big disadvantage if you had to do it. We played with the idea of making this big project that several people might be working on. Maybe it will be a product that can be on. Who knows what's going on with AI. But think that not having access to its database environment both in the security aspect and just developing, adjusting things, removing it a rebasing, doing things. directly within the of I personally that's very important. Jacob Kroon: You buy it straight down because you notice very big limitations if you want to... Some functions, I can't really finger on which ones are, but I if they are webhooks or endpoints for some functions that I want to like email. I want to a function where you can send an email that automatically goes into the system and then I an endpoint and I can't get out. I can't prompt myself to it. Then need a superbase user that you can go into and handle yourself. ⁓ Sebastian: Hmm. Jens: But just Loverball, they have integration with Reason, So they be to build that in. They probably the before, we might not be able to into Super Bass to do this, but you can use another connection to solve the problem. You get bit confused as well, because these things work in so many different ways. Loverball does his thing. Jacob Kroon: and score a prompt that they're from. ⁓ exactly. how expect. Jens: Windsurfing, I'm into jazz music stuff, but you Sebastian are into cloud code. You get really into how things work, how your flow is, how you work. Jacob Kroon: Mm. I will probably always use Laveable to a certain extent, but I have now over to Claude Code just to learn with it. And in the beginning with integration with SuperBase and to understand how it works was a damn hassle. It took time, it was frustrating. Sebastian: You Jacob Kroon: Now I I have it in place, nice and have a good understanding. I haven't done that many projects yet, but I started with an iOS app now. Jens: Mmm, cool! Jacob Kroon: That was tough. It took time in now it really well. I just didn't understand that this TestFlight Expo user that I have has limited amount pushings in a month. So I pushed everything I did. Directly to the test app that I on my phone. So now I wait until June 1st when I can continue. Jens: God damn it. Sebastian: Start on there. Jacob Kroon: ⁓ ⁓ ⁓ Sebastian: They're called the shift map. Jens: Have you seen that they have... Now I'm to talk something else, but it's an interesting anecdote. Anthropic went out and started publishing different types of payment sites now to see if they could increase their price by about 10%. So depending on whether you a or not and which country you came from, got different types of shops to measure conversion rates. So it started to go away. Sebastian: Yeah, they're about that. Jens: But... But, it's a very interesting thing. Do you know why the is so damn expensive? Why it has become so? Do you know the of that? I yes, it is used in AI, but do you know how it used AI? You use frames that... Think of a computer. All your files are on your hard drive. Jacob Kroon: I would say that it is. Jens: But in AI, in to be respond this quickly and collect data this quickly, store all data directly in the Because that is the memory, that is, the closest processor, closest brain. And is so much data, so many complex things that are to so you use RAM as storage. Sebastian: Ooooo ⁓ my god. Jens: Exactly. can understand how the data centers How much RAM is in there. It's crazy. And then you also a compression. Imagine you're You have a picture and you type on that picture. I say, you store that the with the coordinates we said in vector database. It should have... It ends at 1000. Then it into the SSD and checks, okay, but we're not to 1000. Jacob Kroon: Yeah, sure. Sebastian: Massive. Jens: How low is And there is the full-scale file or whatever you want. But because it's so massive, even when it's needs so much RAM. Jacob Kroon: you Jens: is that on the net? Sebastian: That's crazy. But you would have to find a new technology. new type Either we just that type of to all incoming computers and phones and all other types of electronics. That will take the new thing. Or we to find a new way for AI to work with this when we build that infrastructure. Because it feels problematic Jens: Loggings too. Jacob Kroon: How does it on How does it on that market? Is there someone who has attacked it? That's the Jens: I say that this research, been some new artists that are getting closer to using SSD disks with similar performance as Ram. So the industry's direction, as I understand it, is working towards being use normal storage types for this type of retrieval. And that's... Compilation protocols and different search optimization structures. are two big index algorithms that are competing today. It's HNSW versus EVF Flat. If they're so nice names. ⁓ HNSW is, imagine, a spindle network of motorways. Sebastian: Yes. Jens: really big, millions of cars are ready. You can take it from city to city in milliseconds. It's like layer 1. Then layer 2, you're in the city and driving on the small roads. And then layer 3, you're in a few addresses. That's how HN-SW is built. But it also makes that there traffic here all the time. There are lot of connections that should be active and even. And there is 100 % memory that is in IVF flat it's more like think Sweden and then we out hundreds of nodes on the map and that's nodes or it's post addresses, post codes works really well in this a later and every document or that we load into this gets its own postcode and then we know that these have postcode X then we just to up the postcode and search there and then we don't need to use as much RAM. The previous drawbacks with both inVF flat are compromised and much slower in the drawback, but can handle enterprise amount of data at a reasonable price. which HNSV can't do. It would be bankruptcy if you go in with too much data. So if I were to sum it up, database structure with RAG, vector databases, knowledge graphs, index regulation algorithms such as HN, SV and IV flat, ORM tools that we didn't even mention the of, like Bristler and Prisma, which interpret code and data in the Sebastian: Okay Jens: And then the top layer is a langchain or example llama, whatever it's called, orchestrates how the retrieving is to Yes. Sebastian: Yes, but how much do I the list, Jakob? Can you say that you are in everything? Hahaha, yeah. Yeah, that's right. You're Jens: As you can it's very complex. I'm new to this as well. I've it for too many hours in evenings. I find it very interesting. And as I mentioned in the it's necessary pain to know how optimize your own dig. But this is the next level of knowledge. It's difficult. And also, you think about the itself, it's going so fast. Everything! This is also so fast. If we have this in two, three, four months, we'll probably be talking about other brands. And they've up with some new indexing algorithm. HNSV, it's not even It's...unfair. What did you Sebastian: Adé, probi mot pe spajce! I'm sorry, okay? Jens: They're not even there, would say. The big enterprise companies haven't even started working large numbers with AI because they have so much data and things to look and guidelines and things they have to follow. Mm. They probably with data in that way, but here we talk about how we should have a well-trained model. I'll mention it in the that an LLM, an AI, is trained on billions of scenarios, text and data. It's its memory, its personality, soul, who it is. Here talk about how we take a person's new choices, like... to be to to things and make it as easy for those who can relate to existing structures. Training an LLM on data, is, Antropiq and OpenAI and so on, is a different process than learning or taking what they have done and getting it. I would say I'm also cheap. Sebastian: But okay, we're gonna have to end this now guys. to wrap this I think it would fun to you some more episodes, Jens. It could probably be more than one. It would be fun to see, Jackie, how is it for you in your series, where are you how have you built the database? Maybe an episode like, if you just want to get started with your own development project, what should you think about? What can you it for and why? ⁓ Some general tips that you've learned on your own level, like you Jens have learned, like you Jacob have learned, everyone has different experiences, so it can be good and fun to Jens: Absolutely. Sebastian: Yes, that's right. I wonder how for those who listen this. ⁓ Jens: Mm-hmm, that's super cool. Island of Oxlare. I give a shout out. I like the world crying, but it great to have him! Sebastian: If you to Jens gesticulate about the and distance, can watch the video on Spotify. Good Arne, ⁓ Jens: Yeah, hi! Sebastian: Hey, boy.