From LLM Knowledge Bases to LLM Artifacts

追加: 2026-05-22

[00:00:23]Hello everyone.

[00:00:29]>> Hey Elvis.

[00:00:31]>> Hi. Hi Chris.

[00:00:33]All right. Um, >> hey everybody.

[00:00:36]>> Hey. Hey. How's everyone doing?

[00:00:41]>> This is This is super timely.

[00:00:45]>> Yeah. Yeah. Super excited. Super excited for today. I have so much to to showcase today. It's been a couple of months just um you know with my head down just trying to experiment a lot with AI. So there's just a lot to share. Um so happy to see some >> Is this going to be recorded by the way?

[00:01:03]>> It's going to be recorded. Yeah, it is.

[00:01:05]>> Yeah, I think it's recording already.

[00:01:06]Yeah.

[00:01:08]>> So let me just All right. Just set up everything here.

[00:01:15]Yeah. I'm going to have my chat here.

[00:01:17]Right.

[00:01:22]Cool.

[00:01:25]Yeah. Yeah. So, I see a lot of people joining.

[00:01:29]All right. Feel free to introduce yourself in the chat. Uh that was good to um yeah to see everyone here.

[00:01:39]Okay.

[00:01:40]All right.

[00:01:43]All right. Let me just share my screen and get started.

[00:01:49]just share.

[00:01:53]Hey, let me know if you can see my screen.

[00:02:03]>> Yeah.

[00:02:03]>> Yeah. All right. Awesome.

[00:02:07]Cool.

[00:02:08]All right. Great. So, we're going to um I'm going to get started here. Um it's been a while since we have done one of these. Um we've been pretty busy behind the scenes trying to um really take advantage of some of the um you know newer capabilities of of these systems these AI agents. Um we've been experimenting also like what the future of learning is going to look like. Um trying to integrate a little bit more of like the agent stuff into all the stuff that we've been doing in the academy. Um I'm also as as some of you may know like I think community is really important these days like staying connected with people. Um sharing you know I I we want to do more sharing. We want to do a lots of these like build sessions which is something we're going to um be doing in the next couple of weeks. Um as we build out the platform there's going to be a lot of stuff um and a lot of tools that you can use um to really skill up yourself. Right. And this will allow us like our platform will allow us to to really um showcase what's possible with TI. So I'm very excited to share a lot of the stuff that we're working on today um and some of the stuff that's coming down the pipeline. So um so I'm going to do a quick intro for those of you that are very new here. Um so my name is Elvis. I'm the founder of dei. Um basically what we do is we do like I'm an independent researcher behind the scenes. Um I'm also you know some somewhat of an AI engineer as well. Um do a lot of work with companies do a lot of partnerships help companies like you know build products around AI agents and so on. Been in the space working a lot with um you know startups like a lot of bigname startups um and so I've been trying to like bring some of that knowledge and share it in the best way that I can in whatever form I could and capacity um using this academy. Right.

[00:03:56]So I'm also the founder of the academy.

[00:03:58]Um and we have so much down the pipeline that we want to do. Like I think very uh deeply about the future of learning. Um that's something that I'm passionate about and it's something that I think won't change for me. So this is a great technology I think that we want to like take advantage of and you get you you'll get to see today you know how we're trying to how we're thinking about this and how we want to integrate agents more into the learning process. All right. Um so the first thing I want to I want to say um so we have this community right I mentioned the the importance of community like I would like to see people you know sharing talk more about what what kind of challenges they have what they don't understand about the space and so on um and I want to encourage more of that so that's the first thing I want to highlight here the community um and so um you know introduce yourself that that'll be great just to get things started um and we also created this this is like a channel. I'll I'll share this channel later in the chat. Um, it's a new channel here. I want to focus more on like, you know, people that are building with this technology. So, introduce yourself there. Um, I'm going to be posting something there in a minute um about something that we have been kind of working on behind the scenes and we're going to be sharing in in in a minute or two. Um, so yeah, that's more about the community. So, we have a lot of events that we want to do. I think for me um the most effective way to teach about this stuff right we we do courses we do workshops there's lots of recordings here and there that we share uh but I think this live connection is really important to like get to share right what people um are experiencing what are the challenges how people are thinking about the future what they're building and so on um so the live sessions we'll do a lot of these um I have a whole series of them that we want to do which are going to be focused on like building right so how do you build with the stuff how do you do evals how do you do all of this really cool stuff.

[00:05:48]And I think we have built the platform in a way now that I'm very confident that we can do a lot of this now and we can kind of incorporate some of that knowledge again that we are learning with our interactions with with companies that we work with um on AI engineering and and research. All right.

[00:06:02]So this is the one that we're going to do today. Um we're going to be focused on LLM wikis to LLM artifacts. Um so you saw a lot about artifacts. There's been a lot of sharing here and there, lots of opinions about why are these like interesting ways on how you interact with agents and how you can learn from agents. Um, but how do you actually utilize the outputs of agents? Because I think the issue with agents today is like you can you can ship code, you can do a lot of different things um with with agents. Um, but I think the thing that we're not really thinking deeply about is like the output, right? How do we how do we take advantage of the output of the language model or the agent and how can we best utilize that to make better decisions, right? whether it's about shipping code, whether it's about research, whatever that may be.

[00:06:43]And I think that um that requires a bit of like um you know thinking deeply about how to how to present what the agent is offering and also like what's your role in that in in in in working with that agent and collaborating with it right so this is why I got excited to talk about this stuff even though we've been working on this stuff from like last year. Um, so we did one session I want to call out here. I think we did a couple of weeks ago. We have a lot of people interested in this, which is how do you build a knowledge base with agents and how do you do it in an agent quick? Um, I think there's a lot of stuff to learn here. I think it's very early there. There's like like I said in the in the session, you can watch the recording. Um, there's so many different ways how you can kind of build this stuff. Information retrieval systems, right? How do you search um, you know, large scale information and that unlocks a different experiences and applications with language models and agents. So if you want uh you know want to understand what it is, what is like an LLM wiki or knowledge base um you can check out that recording. Um today we're going to move you know to the next step right the next step is okay we have this abundance of information knowledge maybe that we're creating in our companies or you know for maybe personal consumption. Um but how do we how do we present that right the question is how do we present that to a user so that you know like an expert or someone that's working with the agent and collaborating with the agent how do you present that in in a in a way that is more actionable right that you can take action in it and you can get real value from from agents so this is kind of the question that I've been thinking about and I think HTML artifacts LLM artifacts I call them um is a good way to to uh take advantage of agents so let me show you what that is um in a minute I'll show you that but that's kind of the premise you know behind the uh the session today. Um so that's a brief intro. I would say um we've been thinking hard about the problem. I think uh you know everything is going to be opinionated. Everybody has their own different use cases for how they want to unlock agents. Uh but I just think like it's a good tool to like just continually learn, right? Keep updated with stuff. Like that's how I've been using it. Do a lot of research stuff and so on and keep up to date with things, right? A lot of I get a lot of questions. How do you actually keep up with research everything that's coming in, right? all the different approaches, methods on how do you, you know, how do you do evals, all of that stuff. How do you actually keep up with this thing? Uh it's impossible like there's so many so much information, so much noise. So I've been using agents to kind of filter that information and present it in a way that I again can make a decision, can take action on it, right? Whether it's like helping a company on some research project, whatever that may be. Um so so that's kind of the main application that where I'm using these artifacts. So, I'll show you some examples of that. But I've also been thinking like how do you use this um you know at a personal level, right? How do you personalize outputs of an LLM agent and and that unlocks things like different learning experiences um you know unlocks things like maybe building applications, building services, building different things um for your community, whatever that may be. Um so there's you know it's a very I think um there's lots of opportunities what I'm trying to say. Um so so those are kind of again that's the premise behind the session today and I'll get to some examples in a minute and some demos. Um so there are two different releases that we have made to help on on this front right and two of them is and these are like very experimental like they're beta features go play around with them any feedback that you have I'll appreciate any feedback right the more we get people to try it out I think um you know I think all of this is going to be very new to people but I really want to encourage people to go and try it out. So I have one here which is admin only won't be available to anyone yet but this is something I'm working on and you know this one be quite extensive because I'm going deeper into how do you actually build knowledge bases with agents but I want also the value to be like how do you actually do in a company in a in a company format um so it's a little bit more on the enterprise side but I think it has like personal application as well um so that that will come soon um and again that will not be the focus today it's going to be more on the HTML artifacts so all of this will complement um you know each So there there are two labs here that we encourage people to try out. So we we we did this talk I think a month or two ago on you know how do you use agents incorporate agents into like an image generation process for instance that where you need more skills maybe you're not that great at generating images um with agents how do you actually do that with agents um and I kind of build a skill and and and walk through that process there um same thing if you are not familiar with skills I think how do you build an agent skill is quite an important skill today um so go learn about that here it's you know what I've built here is something where you can interact interact with it. You know, you can read. I think this is a little bit like the code academy experience. You can build. You have a little uh sandbox here that you can interact. You have a couple of checkpoints that you can go through. So, you get an understanding of what is a skill, right? What how is it structure? Why is it useful? Go through some examples. And I think that's the best way to learn. The best way to learn, I think, in my opinion is just to build stuff, right? Try stuff. So, this is why we build these sandboxes for you guys to like go and try things out. Um again, this will be available for um our pro subscribers only. Um but yeah, go try it out again. If you have any feedback, just let us know. Um we'll constantly be improving this stuff. Um let me go back here. So that's one. I I think this one we literally uh shipped this yesterday. But the idea is like after you learn this stuff, you know, you go through the interactions, you go through the examples, what's next, right? What do you do with it? I think there are two big things that people are asking me a lot, right? So how do we actually build with this knowledge? How do you build with these tools, these skills and so on? How do you actually build that? Right? How do you put that into into sort of like a workflow and build something useful um you know for your workplace, personal life, whatever that may be. Um so that's the question we want to answer and we want to get people encourage people to build stuff.

[00:12:24]It's not that hard to do it with agents.

[00:12:25]But I think what we need is we need a really good environment um that encourages that and that uh you know like helps people to get there right to that to that um to that place where they feel confident and they feel like it's a useful experience to be building with agents. So this is kind of our answer to that. You know again very early. I don't I think this will evolve. I don't know if it we will keep this but it you know it depends on the feedback. It depends of people if people are actually interested in it using it. Um but basically what you see here is a way to like build artifacts right we've been talking about artifacts. So let me just give you a little summary on what you know how do how do we define these HTML artifacts and why they're useful. Um, and for that what I'm going to do is before I build anything, I'm going to go back here just a little bit of like a recap of how we went, you know, to this, right? How we came from like a couple of months to this to this place and and why people are so excited about to talk about this stuff. Um, so I did this post here, you know, went viral, but this is this is like I I I quote tweeted, it took me like a couple of minutes to actually quote tweet Karpati here. Um but he he he mentioned this thing about LLM knowledge bases which is something if you have been around in the academy for some time this is something we've been discussing from like last year um you know I saw I saw the opportunity to like use LLM agents to like expand knowledge right I am very obsessed about using agents to like do knowledge discovery because of my research background and that's kind of how I'm I'm using these agents right um I've also worked in you know projects like papers code that accumulates all of this research knowledge and one of the things that we wanted to do um was like make things more actionable, right? So when you see a paper, when you see a result, what what about that result is interesting? What how can I expand on that uh particular result, right? What new interesting research questions can we can we ask, right? And we can use agents, you know, to collaborate on this. So this is kind of in the future, I would say. Um and then this is connected to this idea of like self-improving like recursive self-improving agents because you know once you can give the agent proper research questions I think it is possible maybe in the future you know this is what I'm kind of predicting it is possible then to use things like outer research um and do this kind of like loop where you can you know do this kind of knowledge discovery stuff right you can you can tune agents you can tune language models to do to do this discovery So all of these ideas are sort of connected. They're all over the place I think. Um and so we're trying to kind of you know put it together in a in a nice workflow. Um so okay so so this one knowledge base go read about this stuff is really interesting. Um there's a lot of thoughts about that right I share my own thoughts about this. I was experimenting with this from from April.

[00:14:59]Now the way how I looked at this is like okay you can build a knowledge base right but that's for an agent but most of us are collaborating with an agent.

[00:15:06]We need a way to um to make decisions on that knowledge base as well. We actually are part of the process. I don't think a knowledge base can just be built you know in an automatic way and that's it right like we still have humans interacting with agents as far as I know like um there's not like agent to agent stuff that you know that that that has unlocked um like new discoveries and things like that I I think we still need human expertise and human verification is really important here so I I asked myself that question right and I think this is how I kind of came to the conclusion that we need different ways and forms on how we um interact with information how we present information, right, that's outputed by these agents.

[00:15:43]So, this was the best example that I could could come up with at the time, and this is something I was exploring, which is I just created this interface.

[00:15:49]Um, I have this ID that I built for myself. Um, and it's all very visual because I'm a very visual person like I I love visuals. Um, maybe because of the work that I do. Um, and you can see here there's a lot of information. There's like results. Uh, there's like papers.

[00:16:01]All of this is backed by an actual knowledge base. Um, and it allows me to keep up on the frontier of like all research. and al it really helps me with the work that I do and it helps me with all my personal projects as well. So I wanted to kind of you know put that into like a little product experience um and share with people how you can start to play around with these ideas and and this is why we built the eye builder. So I'm going to show you in a minute how what what are things that you can do right now and and what's coming down uh down the road where you can actually build things like what you see here.

[00:16:29]Okay, so that's kind of the the motivation behind it. Um and then you know like fast forward we we I showed more examples right how you can consume my YouTube videos there's many examples like that I've been sharing my timeline um if you're interested in that again I'm going to share some links later when the recording comes out um and then like fast forward one month after you can see it here in May um you know T published this article like HTML the unreasonable effectiveness of HTML and I think he came from a different angle um my angle was about information how do you make better decisions with information that you have that's my use case. His angle is how do you actually from a code you know engineer perspective how do you actually take advantage of the outputs of an agent or or a language model um and his his answer to this is not markdown some of you are using markdown right we're all putting markdown in fact the wikis that we showed the last time are actually built on markdown and he you know came to this conclusion that maybe the better way is is to let the agent produce HTML artifacts um and and and this goes deep right because like the examples that he shared like there's a nice little blog that they posted about this. Let me share with you here this link I I'll share with you cuz they posted it in a blog this I mean this went really viral um and people were really really um excited about this stuff. Um again for me I think it you know this is something I've been playing around with. So um you know it was it was not it didn't feel like it was totally new. I I just it kind of validates everything that we were doing.

[00:17:54]Um all righty. Okay. So let me Yeah.

[00:17:57]Okay. I shared that one there. Um and then there's this nice little page if you missed it. This I think is a is is the best resource that was published here. Um I like it because there's a lot of examples, right? I mean I'm actually going to try to do some of these examples today and using our AI builder.

[00:18:12]Um but yeah, I mean the comparison between markdown files and HTML files I think makes no sense to me. Like I disagree with that. Like um so so the argument here is that um you know the best output format of a language model is not markdown but it's HTML because it's more consumable. It's visual. It has all these components that a markdown document could not do. Um, it turns out LMS are actually really good at generating things like XML. Markdown obviously they're trained a lot on markdown files. Um, and HTML as well.

[00:18:40]HTML has a structure and all of that stuff, right? So that's kind of what he was kind of mentioning here, you know, and there's a lot of conversation around this, right? Obviously, it's markdown I don't think, you know, can be replaced by HTML. To me, it's more like combining both of them. So I'll show an example of how we can combine both of both markdown and HTML how they can work together. Um so that's coming in a minute. So you can see here there's plenty of examples. Go through these examples. I tell you there's so much good stuff here like this. This is like really amazing I think um artifact that he created. Um like how do you do for how do you build exploration and planning, right?

[00:19:16]Planning is super effective when you work with agents. Um how do you do code review for those of you that are doing engineering stuff, software engineering?

[00:19:22]How do you use design? Right? If you're product, you know, manager, if you're doing design design team, how do you actually use agents? How do you use these artifacts? There's so many good examples here. And all of this stuff, I think, will be possible to reproduce with our um with our own system, right?

[00:19:37]I'll show you an example of that. Um, you know, animations, right? All the visual stuff, illustrations, and and the list goes on and on, like decks, research on learning. Oh, the learning stuff I really love because that's kind of what we're doing. Research and learning is really our um the use case that I focus on a lot. Um, so I really I really like this like explainers, concept explainers. We'll we'll touch on an example on that in a minute. Um, you know, how to generate reports as well.

[00:19:59]You might be in a company, there's so much information coming in. The Slack channels are getting insane. How do you take a lot of that information and then, you know, connect it to an NCP or agent on MCP produce, you know, something that's like really nice, right? Really, um, I think could be quite significant and of high value um, in a company. Um, you know, customer there there's so many examples, right? just go through some of these examples when you have a chance.

[00:20:23]Um like I I like I like the diagram stuff. Um you know I think the coding example was super good. U there was one in particular I really like this one. I believe it's this one right here where because reviewing is hard, right? Like I think we should be reviewing code, right? If you're doing code stuff, if you're doing agents with code, please do review. I encourage more more people to do this. I think um it's actually good and and and these type of like solutions do help with that. Um so I I like that this example was put here u which is telling you you know they do care about reviews even though you might be shipping a lot of code right maxing whatever um it's it's important part of the process right like let's do better than than than sloppy code let's do better than just vibe coding we can we can do better than that and I think these type of things allow and encourage that um so I really like this example this is a fun one I think this is a really useful one and there are many ways you can do this um anyway so those are kind of the uh things that I wanted to share before we get into the actual example. Um, let me see at the chat there was some questions I think. Um, okay. Let me see.

[00:21:24]Yeah, I'll get into the I'll get into the AI builder stuff in a minute.

[00:21:29]Let me see if there's anything else here before I move on. Yeah, let me um knowledge discovery I believe. Okay, thanks AB. Thanks for sharing that. Very useful. Thank you.

[00:21:41]Okay. All right. All right. So, I think that's it. Yeah. Keep the questions coming if if there's anything um any resource here. I'll post all of these links in the um in the recording when the recording comes out. So that will be available for everyone here in the live event stuff. All right, so let's get to some fun examples. So I've been building with this stuff. Again, it's very new.

[00:21:59]Um not perfect. You might see bugs here and there. If there's any bugs, please report it to us. This is super experimental. Um but it's like how do you how do you build something like this, right? How do you actually, you know, these use cases? How do we do this? Because I mean I saw this and I was like maybe I can replicate it. I can maybe uh give my cloud code or my codex and maybe codex can like put something together. But I think there was more behind this and what I realized very quickly is that if you look at the design of all these artifacts they're quite consistent. They're using very similar design components. So that means they're using some kind of skills and so on and that's very hard to replicate. So what I've been trying to do is kind of think about like how do you build like a like a skill like a little sandbox that allows you to kind of build this stuff.

[00:22:40]So, it's not that hard, right? And you're getting good results and you know, you're getting encouraged and you have that confidence. And this is what we've been doing with the the eyebuilder. So, let me show you in a minute what that is. Okay, here it is.

[00:22:50]Um, so what I'll do is um there are a couple of examples that I can work on, but I I want to show you um something that I've already built. Um and then we're going to work on on on a few examples here. Again, this is only going to be available for people on the pro subscription. Um, by the way, this whole thing here, um, we're using we're using the the the cloud models, right? So, we're using this LLM and they're quite expensive, I would say. Um, so we need to figure out like, you know, how much what's the what's the the rate limits and all of that stuff. I I kind of need to figure out that that out with my team. So, again, very experimental. If you want access to all of this stuff, just let me know and and we can work something out together. Okay? If you're hitting limits, reach out to me. I'm happy to to to give you more um usage so that you can experiment with this and and we love some feedback on this stuff.

[00:23:35]Um okay, so we'll load this one here. Um so this one is is like I have three like examples that I just kind of quickly built. Um so this one is like it's basically the idea of the wiki. So we built this little wiki here and it actually uses this um this wiki builder skill that we built and we we open sourced this. You can find it in in our GitHub. Um and we just load it here in the sandbox and you can directly use it like you can just go you know select it and you can directly build your own wiki if you want right go with the agent and just tell it to build it for you. So what I told it was build a simple wiki of methods use um from memory in AI agents just simple as that it packages all of the stuff it loads the skills right um it has a front-end design because again these are just going to be um skills that you can use to to better improve the design of this. If we were not using a front-end design, it would probably not look as good as what it looks right here. So, this is what I was saying, right? I think Entropic uses some kind of set of skills, uh, internal set of skills and that's why all the design components look very similar and I wanted to kind of mimic that here and and they open source this front-end design skill, which is what we are trying to use here. Okay. So, we use that in combination with web search and then we just asked the agent, can you just go build like this wiki for us?

[00:24:48]built the wiki, built the the artifact, and in just one prompt, it produced this right here. Okay, so it has like an overview of like what are the different memory um techniques. So you can see here like we're building this um it's giving us some information. You can use it to sort of learn or maybe do some recap, right, as as you develop this. So you can see all kinds of applications for learning here. And this is what I was kind of excited to like integrate this into our platform. Uh but there's so many other use cases that this unlocks, right? So, this is just one of them. So, I'm going to show you another one here. Um, let me see. I think it's this one here. So, this one I think is one of my personal favorite. And if you get a chance to try the AI builder, I would encourage you to actually try this one. So, this one is using I believe it uses the lesson generator. So, this lesson generator basically is a skill that I built. Um, I think this one is not open source. I I I will open source it. I I want people to have access to it anywhere they work. Um, so I will open source it maybe later. um or early next week if I have a chance. Uh but this basically what it does is have like a set of instructions that's going to produce this artifact and this artifact basically is like you can learn any topic right you can ask combined with web search um you can ask it like just produce like I asked it here build me a set of lessons that goes deep into black holes uh for a high school student you know please add engaging quizzes knowledge checks whatever so if you have kids around you know um you can build a personalized lesson for your kid um and it uses this now again very early days.

[00:26:20]But our vision with this is that um you know as a first step we want to make sure that the agent is producing high quality information obviously high quality lessons and courses but from here I want to make this more interactive right like the agent can interact with the with the um artifact.

[00:26:37]The artifact can then also interact with the agent when it needs. So it's all more integrated. So that stuff is coming in the future. I have a like a vision for what that may look like. Um, and I've been playing around with some prototypes for this. I think in the future what we'll see, we'll see more of this stuff. Um, this is not a coincidence. It's not I I I think it's not um, you know, it's not something people just want to hype. I think people resonate with, you know, information when it comes in a specific form. Um, and something actually that Karpi did mention in, you know, as a comment here, it's like look at this part here and this is the thing that captured my imagination and this is something we've been thinking about. Um, so first you have like raw text, right? Then you have like Markdown, but Markdown is very hard to read. I mean, I I prefer Markdown by the way. I I love Markdown. Um, but then like tables, they look really weird and so on. So then there's this HTML stuff, right? And then there's like four, five, six, whatever comes next. I'm sure we're going to unlock something here. But then to me, it's going to be, you know, the future is going to be something that just you ask it and it's just presented to you, right? um with all the multimodal capabilities of the system you can produce simulations you know you can run evaluations you can do research of that's really I think the frontier that's really I think in the future and there's a lot of stuff happening in this space already so I think it's just we're taking these next steps right all of us should be thinking about this it's not just about learning right personally but it's about you know how can you make best use of what this intelligence system is producing that's kind of how I think about this right that's what has inspired me to really think about what are the tools that we're going to you know provide for like our community right so our community I think should have at least minimum be able to generate things and generate interesting information that's high quality so yes in combination with the front-end design skill um we build a pretty like nice little in like interactive thing here and you can start learning about this stuff you can see it goes through some topics right um this is really fun I actually used this u on some topics um like I wanted to I I wanted to learn a little bit more about what you know people are working on in terms of self-improving ages. That's a topic that I'm going deep into as an independent researcher. Like I wanted to learn about that. Um and I can do that. Let me show you like really quickly how I can do that. So I'll just go here and then I can go um and select right I want to do some a bit of web search. I'm going to enable that. I will leave the front end design is a default because again if you want pretty artifacts I think this is good to enable. Um and then what I'll do is I'll just select lesson generator. So that's selected and then I want to generate, you know, a set of lessons. I want to I want to see what's up with like self-improving agents. Um, so I'll just do like a little prompt here. Um, again, you don't need to do much and you already get good results because it's a sandbox. It's already packaging all the instructions and everything for you to build something useful. Um, and before I do that, um, all of this stuff that you see here will be fully integrated eventually with all of the content that we're producing, right? So you'll be able to access all our courses right our lessons our labs our code skills everything that we share will be accessible to you here. So this is going to be you know maybe this is not like the end product but this is going to be something that we are going to fully integrate eventually into the entire platform so that then you have something that you can interact with and you can learn in whatever you know make it more personalized at least experience right that's the vision here. So let me just build this thing for you here. Um and then you can try it later if you have access to it. So um there there are two things I want to do, right? So I want to I want to do um like first of all I could produce I could produce um like just a lesson set of lessons. So I'm going to say um can you produce a set of lessons on the topic of uh self improving agents?

[00:30:33]I'm going to scope this a bit. I'm going to tell it uh just want to learn the applications right applications and the latest research and methods around it. It'll be good. Okay, let me just it will be good. It will also include um results uh interesting results results on the topic, right? Like there's no there's no limit, right? Like you can you can ask it anything. It'll just go on the internet, search for whatever, you know, um create the content, then present it to you beautifully using the front end design skill. That's it. Like you can do a lot of the stuff. Now, keep in mind we only have a couple of skills here. Um, I'm going to get to the wiki builder stuff.

[00:31:22]That's really that's kind of what this talk was about. Um, but you know, like we will be be adding more skills, more tools. You you can you'll be able to do a lot more stuff with this, but it's already quite powerful, right? So, you can just go ahead build that. Um, and then this sandbox, what it does, it's going to load a coding agent behind the scenes. Um, I'm not going to get into the details of what the coding agent is.

[00:31:43]If you're interested in that, we we probably can do a session on that later on. Um, but yeah, it it's just going to go ahead and and and just build this, right? It's going to open up the workspace, load the sandbox, take all the skills, use all the tools, all the system prompt instructions, everything, and it's just going to build the stuff for you. Um, that's kind of the experience that I wanted to deliver with this. Um, you know, later on you'll be able to like generate code, right? Be able to manage the code and all of that stuff. Um, we're not doing that yet. I think that's a bigger scope. Um, but yeah, we we we again we want to encourage building. We want to make this as as fun and as useful as possible and so we're just taking uh steps to encourage people to try it out. All right, so let me see what we can have here. So let me see if we have some questions here as this builds. Right, this this will take a minute or two or three. Remember it's a coding agent.

[00:32:31]It's access to the file system. It's doing a lot of stuff, right? It's a sandbox. It's doing a lot of things behind the scenes. So uh here we're just it's just opening updates, but it's doing a lot of things behind the scenes.

[00:32:41]So it'll take a minute or two. All right, let me just answer a question here. There was a really interesting question here on the chat. Um, so again, if you have a question, please leave it in the chat. Um, or if you want to unmute yourself, go ahead and unmute yourself. I'm happy to to take a question from anyone on the chat. All right. So, with the rise of LM wikis where documents are predigested into structure agent friendly JSON before indexing, are we essentially building a better rag or replacing it entirely? Is the retrieval step still necessary if the wiki already captures everything important? Now, I mean, I have opinions about this, right? because I I I do think about information retrieval. I mean, this is I I got a PhD on information retrieval. So, I got to think a lot about information retrieval algorithms like TFIDF, you know, versions of TF, you know, like um BM25, all of these things. So, so I think a lot about like rank systems um and and and I'm very careful because like I come from that world of information retrieval. I might have a a little bit of a bias, right? Like I I I think like BM25 is super powerful, right? um you can use that u for to build search systems. I work at a search company like elastic. So I do understand how powerful that technology is when you ship it correctly when you implement it correctly. Um so you know I come from that world and I and again I might have that bias but I think in the end with agents like things are evolving and changing so rapidly right like like how we used to do how we used to do like image generation how we used to do um you know code generation all of these things are evolving and changing. So we have to have like I think an open mind right and this is something I've learned like you know I was such an advocate let's do rag let's do you know the the TF BM25 let's just do that right but then like eventually I realized like neural search is also very important right because neural search captures things that you know a BM25 algorithm wouldn't be able to capture so in my world it's a combination of the two right so basically what I'm trying to say is that we don't throw away what we have learned from rag We take what we have learned from rag and then we use all of these new things that agents unlock like agentic search is really powerful and this is what you you know I think um um you know um habib is is is talking about and then how do we build a a better hybrid system of the two and then you may think okay well this is you know why am I why do I care about this but actually it does affect um you know it does affect the latency it does affect how how um how capable your application is Because if it's not searching for right information, like I said, right? The agent can produce stuff. It can hallucinate stuff. If it's not able to find information, you know, it won't produce good results, right? No matter in which whatever way you present it, it won't produce that result. So like I think it's a combination of the two. And where I've seen this uh in particular with coding agents, I think with coding uh products like cursor for instance, Windserve and so on, it's very interesting because they have different ways on how they do search, right? Um and I've been trying to follow all of the tech technical stuff that they've been approaching and and to me it's very clear like it's a combination of the two like if you use cursor today they use semantic search they use you know they use this this this these rag solutions um because they know that's really important when the code base gets large the language model and the agentic search at GPS's you know search is not enough it's just not enough right it won't capture the breadth of the search that certain type of task you know in a codebase requires so I think you It's a combination of the two that usually works well. Now then you have things like cloud code where you know there's no semantic search. You can implement semantic search if you want but there's no semantic search and all you do is just like it uses GP or some some you know some like uh something in the file system or something like that um you know some function really really simple function um and it doesn't work as great. So you might be oh but cloud code is not comprehensive right it doesn't it doesn't understand the code as much as a codeex would or a cursor would. Well, that's the the answer is that's the reason why it doesn't because it doesn't do it cannot search to that depth, right? It doesn't mean that aentic search will not improve. It doesn't mean that I think we'll probably figure out a way to make it work. But I think it's like really bad idea to say that oh agentic search really um you know rag is dead and and and so on. I think this as far as I know like my experience in in computer science um everything that I've learned in the history of computer science it's usually going to be a combination of technologies we learn what we can we pull what we what we think is good from one technology and we we pass it to the other it's no different today even though AI is and language models are amazing that's my opinion about it right very long answer but again like I said very passionate about the information ritual stuff um and that's something that I think I want to like dig a little bit deeper because that's just an answer that I gave very opinionated answer but I think we need to be able to show this and the way we show it is through results right the way we show it is like in an application itself how does how does the performance look what what actually does cursor do differently than what a cursor no sorry a cursor or a cloud code would do differently right and I mean all of your experience in this right I see a lot of complaints about people saying oh you know like like cursor is is nice right it has good personality but it doesn't it doesn't understand the codebase um to that dep right and The answer is that the answer is that the search is bad.

[00:37:55]Okay. Uh so let me see. Hold on. Okay.

[00:38:00]>> Yeah. We'll have to watch. Yeah. Anyone?

[00:38:02]Any Yeah, somebody is there. You want to?

[00:38:04]>> Yeah, I thought. Thank you. Thank you, Elvis.

[00:38:06]>> Yeah. Okay, cool.

[00:38:08]>> Yeah, Elvis had a quick question. So, yeah, thank for the session. The session is really useful. So, I wanted to find I was interested in this learners had built this thing that tool that you're demonstrating, right? So how does it like actually like apart from building the like the course details what else does it does it build an artifact for a code that is goes along with this or how does that work? Yeah. So it's basically going to create um you don't see it here but because I selected so front end design is for design stuff right it's just a set of instructions to tell cloud because we're using the cloud models to to like make it a little bit like use nice design principles that's just a set of instructions the and that gets loaded right as a skill um so the lesson generator which I selected here another set of instructions in order for us to get consistently this information that we're seeing Um, so in combination these two would give you exactly what you see here, right? And so you you you can adjust it.

[00:39:04]You can it's basically just a skill file, a set of instructions that we just inject into the agent and then the agent produces um produces HTML, produces JavaScript for the interactions here in the artifact and that's what you're essentially seeing here, right? So it's like like if you like you know like these examples that I shared shared here, this is basically just HTML. You can interact with it. That's it. Right?

[00:39:25]a little bit of JavaScript, HTML, and that's basically the artifact, right?

[00:39:29]There's nothing fancy going on. Now, I think what could be interesting is like how do you make this more expressive and more interactive that, you know, you can interact with it and you can do many things with it, right? Like I think that's you know that this is where we all of us need to be really excited about it because I think there's a lot of opportunities like being being able to present information it has a lot of value like I think this this is already super valuable for me and I'll go through this and I'll take some time going through this I'll learn something about the topic but what it unlocks is like this personalization that I think I haven't really experienced before like agents can produce information but you know it produces it in file sometimes I don't know what it does you know like I said review review you can review code.

[00:40:11]Some of us use the diffing functionality and so on. But with everything else, like we're not seeing it, right? All we see is just the the final thing and we don't we don't get a way to to see what was built in between, right? So >> to me, artifacts are are useful for that. This just happens to be a really good use case for learning and I wanted to use that. But we're going to get into a more fun example where it actually creates because it's a file system, right? It creates files and then it just loads it automatically for you. Um, so, so yeah, that's what's going on behind the scenes.

[00:40:40]>> Yeah, >> I think that's where like I think spec driven development and all of that takes a precedence as well. So, >> yeah, I agree. That's correct. All right, let me see if I have any other questions here. Um, yes, great question. Um, Arash, I think that's a great question. I love this question. Um, so this is specifically for the lesson generator, right? So, I was opinionated with the skill. Um, I left the skill a little bit loose, right? I I just told it like work for any topic basically. But I think that's where I was saying that being able to see what the agent produces and combining that with a little bit of your expertise and whatever you're interested in, I think goes a long way. You can always use like these authoritative like sources. You can use like archive peer reviewview journals. In my case, that would be really important. Um and and I think that's something that us humans, you know, can do, right? Like we're totally capable of doing. I remember when I was doing like my PhD uh back in the day, the first task that I got was like go read papers, go read books. And I remember like that was a that was such a like like like very unique experience that I had at the time. I I hated research. I I didn't want to do research papers like in in in in you know in my undergrad and so on. Um but I learned very quickly the the the use of that, right? So I I think that's where we kind of, you know, complement the agent, so to speak. Like we we we do a little background research. We we help it with the sources, right? Um by the way, like the web search that we're using here already kind of does some of that. So we're using like Exa for search, right, for web search. Um and Exa has these like nice little like um parameters that you can modify here. I don't think you can do it. I don't think it's possible to do it yet. Um but you can modify like what sources you would pick. So you can pick the sources that you want, but I mean you can do it at the prompt level and you can do it at the API level as well. But I think I I love this question because this is this is where I think you can really differentiate, right? So one fun thing that you can do here. So let's say this was like a little competition, a little hackathon. Maybe we do this in the future. This this could be super fun because I think we can learn a lot and we can share a lot from this process um of building these things. So we we can share this, right?

[00:42:46]So I can take this thing here and I can share it. Um well, I don't think it's producing it right now. Uh okay. So there's something weird happening here.

[00:42:54]Let me see. Um, let me go ahead and load this one.

[00:42:59]Yeah. So, so there might be a bug here.

[00:43:01]I'll probably need to check that out.

[00:43:02]Um, but anyway, so like I can produce the artifact. I can I can share it, right? So, I have this link that I can that I can produce. And so, I can load this here, right? And then you can imagine like having a little hackathon where like just producing the artifact, producing the set of lessons um and all of us choose a different way, right? So, so what do you who do you think will win? That's the question, right? Who do you think is going to produce the best artifact? I think it's going to be like in those days where I was learning like, okay, you need to go read about these topics. You need to go and read a ton of papers in that particular topic. I think that person is probably going to have a better chance to produce a better set of lessons for this specific technical topic. So, I think that domain expertise, that human verification, I think we're we're sleeping on that. I think that's just this is why I encourage people to build with it like build with the technology because you can leverage that domain expertise in so many interesting ways. Um so yeah maybe in the future we do something like this.

[00:44:00]I'll I'll be curious to run it just like as an experiment to see who who who like actually ultimately produces the best and we can make it more sharable in in the community. But um yeah, that's kind of my opinion about it. Um you know like I I think I think this is still very unsolved. Um but yeah, one thing we're going to do um as we continue to work on this lesson generator stuff because we really want to integrate it into our platform. Um like we're going to we're going to use Rag here. So we're going to be able to use our transcriptions. We're going to be able to generate lessons based on the data that we have produced based on our own courses that you know we have rigorously went through the material tested things right they work you know and so on run the evals and so on. all the methods that we talk about we we do some eval work and so on and so you know that knowledge is there and and we just want to tap into that and then produce better artifacts and lessons on that so yeah it's it's a it's a topic I'm passionate about but I think the overall point is that you know domain expertise as far as I can tell doesn't go away that cannot be replaced as it stands today maybe in the future something changes but I'm not seeing that today that's kind of my opinion about it uh okay so does this take input from local knowledge base or maybe an API that gives output that I can use into info that's relevant. Um yeah. So, so right now it's not using this this particular example is not using um a knowledge base per se. Um this was just gener is basically just generate well not this one. I think the one that I generated which I completely lost here.

[00:45:29]It might be a little bug that I have. Um I think I know what it is. But anyways, like this one was generated just using um the lesson generator. U so this this is just an HTML file. That's what it is.

[00:45:40]But I like this this question because in fact what I would prefer to do here as I build this out is I would prefer to have like a knowledge base or a little database of sorts. If what I plan to do is maintain this over time I think this is something that you know if you're a researcher you think about a lot right you have all of these papers you you have this database of stuff you want to maintain it um and it keeps growing keeps growing. How do you actually maintain something like this? Well a knowledge base is going to help you with that. Um, so this example only uses HTML, but this is where I kind of disagreed with with with um with Tariq from Entropic in the sense that I think the comparison with Markdong is not a great comparison. And in fact, I mean, that was the title, right? The the the the tweet was um that the new Markdown is HTML. I completely disagree with that because I think we're not really thinking deeply about um how to again combine these things like in the search problem. How do you combine these things? I think that's where um the win will always be is usually a hybrid of these technologies. So let me let me like stop talking there and I'll show you an example of uh something that you can do that's powered by a little knowledge base that's generated by the language model. Okay. So this one Greg is using um Greg is asking about which model. So this is using the lowest the lowest um um model. So it's going to be Haiku, right? So he's using Haiku I believe 4.5. Um so it's going it's using the cheapest one because this as you can imagine is quite expensive, right? like one run of these could, you know, could run you like a dollar or something like that. So I try to make it as cheap as possible to give you like something to play around with um and to get some feedback from the community. So So for now it's using a lower tier model.

[00:47:20]Hopefully the models get cheaper.

[00:47:22]Hopefully I find a way to make it uh use set or something like that. Um and eventually to use a more powerful model, right? I mean cloud models are extremely expensive. I don't need to say much about that but you know most of you will know that.

[00:47:33]>> Um >> Elvis, one quick question. So how did you white code this one? And I want I'm interested to see that one aspect as well.

[00:47:41]>> Which one? Which which >> the learn as you build the tool that you created?

[00:47:44]>> Oh, the tool the the tool itself you mean?

[00:47:46]>> Yeah. Yeah. Yeah. Yeah.

[00:47:48]>> Okay. Yeah. Um we could like I said um wasn't I didn't want to talk about the actual um the actual like bill here of what I did here. But um I think it'll be a great um like we we I did this live training I think like three months ago um or four months ago. I think in December we did it um um you know December to January and and in that live training we're doing cloud code and we were teaching um something that we did with the final project was how do you build how do you build like an agentic product an agentic AI product right I think I might want to bring back something like that because um I think we were too early again um with most of this stuff we are very early but we wanted to show how do you build like an AI product um using this idea of sandboxes right so this is one product by the way that's built on sandboxes Um and everyone is talking about sandboxes, right? Manage agent sandboxes. How do you actually I think most of it is going to I think I was right on that in December when I was talking about most of it is going to go there. All the companies the the infrastructure companies like Cloudfare everyone is talking about sandboxes. Um this is what sandboxes actually allow us to do to build these kind of personalized sandboxes, personalized experiences for for for us to kind of experiment with and and and and use and and build value on top of this incredible technology like agents and LLMs. But yes, I think it's going to be a great session to do. If there's interest in that, go to the community and let me know specifically what you would like to see because there's so many parts to this like mold selection, experimentation, right? Um there's there's so much to it. Like this is why I didn't want to get into it. Um but yeah, let me know let me know in the community if there's interest. We can definitely do something. Yeah.

[00:49:24]>> Okay. So, let me see.

[00:49:28]Oh, I love this. So, so Greg is asking um uh so Greg Greg is asking I'll I'll do I'll answer this last question and then we're going to move to the last example. Um please stay if you can for the next I would say maybe 15 minutes u because I want to show you the best part out of this um out of this tool right like this is where I got super super excited. I got like I I think this is part of the future. This is where everything is going to go um when you include automation into it. So I have an example that I want to show you in a bit that's going to combine everything and it's going to include this automation part. So let me ask answer the question here. Do you have plans to also include generating lessons into remote or hey gen videos by the skills they provide?

[00:50:10]So the quick answer is um not yet because it's already quite expensive to like you know to to to support the sandbox with the model itself and the coding agent. Um, and so what we need to do is we need to first let you guys try it out, you know, see what it can do.

[00:50:29]Um, and and and give us that feedback and we need we need to be able to understand what's what's the cost related because the more you go into like generating videos, images as well.

[00:50:40]I'm I'm interested in incorporating image generation capabilities like the nano banana stuff, I think it would be great to see like a like a like a really cool picture of like a black hole or whatever concept that is and explains it. this is something that I already have built and I can tell you it's extremely expensive. Um, so, so I think like I need to figure out a way with our team how do we actually enable all of that experience, right? And and and allow you to build like more futuristic type of like interactive like artifacts like I think that's a great idea like the motion stuff. Um, we have been experimenting with motion hen videos like 3D, you know, the the world model stuff, the 3D stuff, all of that stuff.

[00:51:20]I think it's going to play like an important role here. Even if we don't do it in the next couple of months, we're going to be talking about that and I'll I'll try to do some build sessions on this because I think this is something that you know like I I think this is a future and we should be building with it. It's not just for learning. I think it's going to be useful for like I said the artifact is useful for making decisions um in the real world and so on. So yes, there's a lot of stuff coming up. Um I'm very excited about that one. Thank you for bringing it up.

[00:51:46]Uh let me see. Yeah, and we have some skills also that we're going to open source on this one too. Um let me see.

[00:51:53]Okay. Anyway, so I'm going to move um I'm not going to take any more questions because I want to show you one more example here. Um and this one is going to use all of them. So it's going to use this one. It's going to uh sorry, it's not going to I don't think that I need the lesson generator. I'm going to use the wiki builder. I'm going to use the web search. And then what I'll do is I'll I'll build this wiki, right? So this will take a little bit more time um compared to the other one. But what I want I'm interested in this, you know, I'll learn a little bit about the self-improving agents. Um and as a researcher I want to understand and by the way when I say research I think all of us should be doing research like this is kind of I I always tell people like one of the the skills that is very hard like you know in in school we learn this like how to do research how to find topics stay up to date with stuff and I do it for a living like I actually do research you know a as as as a profession um but I found the more I talk to companies that want to stay on the frontier like their engineers are not having more time to to do research like they're v coding apps they're doing the review process and they get to stay up to date with models they get to test with you experiment with models they get to like read papers and so on so that's an exciting conversation that I did not expect to see in the industry like um in in some professions but like even product managers are doing it so so I wanted to talk about this stuff because all of this that we've been talking about is kind of coming together and it's showing that people are going to be able to like learn and be able to like skill up themselves and doing research which I think is is is going to be an important um like skill as as as as the technology evolves. So I want to show you like how how that how you can use that here. Um so what I'll do is I'll I'll create a wiki and I'll say um create uh let me see here create a wiki. Simple, right? Like I don't need to be too specific. I can be specific here but I I just want to keep it really really simple. I'm going to create a wiki um using the I don't need to mention it because I'm already you know it's going to be injected as you can see here but I'm just going to mention it there reinforcing it uh scale and um that collects information about the latest research around self improving agents. Okay. So I I I think self-improving agents is the frontier. Basically, the next phase is going to be how does AI improve AI itself. Okay, that's a very research heavy. Um, so yeah, this is what I want to do here and I'm just going to hit build. Um, anything that I want to modify, I can always modify it. So, so there's a couple of things that's going to happen here, right? The the the skills are going to be loaded. Front end design is going to be part of the design of the prototype. So, this one always is going to build an HTML artifact. That's that's by default. Um, and it's going to produce Markdown files. I'm going to show you the structure of it. Um, it's going to use this wiki builder skill that we open source. So, you can find it. Let me see if I can find it here.

[00:54:46]Um, GitHub.

[00:54:48]Yeah, I can share the link. If you want to try it out with cloud code, whatever you want to try it out, if you want to try it out um yourself. So, let me go ahead and find that. Um, I think I put it here. I need a better way to to share this stuff.

[00:55:05]Um, let me see here if you want to go and play around with it. There's a couple of skills and in fact most of them we are actually already integrating. So this one is going to be a fun one. I I I think this is the LLM console if you got a a chance to try is super super interesting. Um, but anyways it's like image generator. We're using that in in the labs. Survey generator is another one that we did before I think in the previous session. And then we have this wiki builder. So this wiki builder basically what it does um um it packages all the instructions for you to to build a little wiki right um on this idea that Karpati was talking about LM knowledge bases um so it packages all of that really nicely um and you can use this in cloud code you can use this in codics wherever you're you're working um and basically has like set of instructions it has like a nice little layout that you can use like it has a configuration of the wiki um and then it has like prompts again if you want to review this stuff we did a recording on that you can find that in the in the platform under live events. There's a recording for this. This is the previous session that we did um where I go a little bit deeper into this and and then it builds like files, right? Like all of it is Markdown files. And then the fun part about this is that the agent can just search around this, right? So Habib asked the question like is this going to replace rag? Well, you know, like it's hard because like I use Markdown, but then my solution actually uses Markdown and then I index the Markdown files. So I actually have the raw, I have the Markdown files and then I have an index on top of that. So then now when I ask a question I have this little router that routes either the the the search to an agentic search which is just searching files or it uses this semantic search like this is so much like this is powerful like when you combine these things you realize like the experiences that they unlock. Um if there is interest for this I I can definitely go deeper into this because I have so much opinions about this stuff and and and ways that I've been building with this stuff. Um but anyway so so at a high level this is just markham files right like markham files where it's just concepts and so on um the the markham files have like you know the front matter stuff um you know have structure everything is in there right so that's that this is what this uh skill file is doing I think the the actual skill file is this one and then it has like references it has examples go play around with this stuff you can use it wherever you want you don't have to use it in our platform like this is this stuff is like is going to be extremely valuable for companies to have things like this, you know, like okay, this is a opinionated implementation of the idea, but go and play around with it. Go try to see if this actually works. If you need to build some semantic search on top of it, how can you use this in your personal life? Right? I'm using this for like nutrition papers. I'm in so so much into the health stuff. I built this like nutrition wiki. I have this research wiki that I built. Um, everything is a wiki. My agent sessions themselves are a wiki because sometimes I see like my agent is like um you know like don't understand, it doesn't have memory. I use it as a memory stuff too.

[00:57:52]Like you can do so many interesting things with this stuff. Has so much applications. I think this is why Karpati stuff went viral is because people were already thinking about this stuff. But anyways, it has like templates, has scripts, has a lot of really cool examples. I'm going to be working on this a little bit more and improving it. Um so that people can really take advantage of this stuff. And I I'll also build something that's has this semantic search using some open source tools to do semantic search on top of it because I think that's that's a great unlocker, right? Like when you can do semantic search on your data. Um if you if you work with like like high density information like research papers like search is going to be the winner there. Yeah. Anyway, so that's the wiki builder. Let me go back here. Okay. So there we go. So we have um I don't like the colors and other stuff but anyways like it tried. I mean it's haiku. How much can haiku actually give you? Like nobody's using haiku probably in this group. But anyways so just let's just just understand that right. Um so we have we have like an overview. We have a little framework here approaches. I did mention something with results and benchmarks like I really want okay you know the common one S swb bench um I got a question from someone in the community like what are the benchmarks that people are using these days for like frontier agent research well here are some really good ones style bench is going to be an important one um and then I can just continue to build on this wiki so what is the wiki let me share an example with you um so let me share what the structure look like so I'll ask it here what does the structure look of the wiki Yeah, it can have a conversation. Um, and you can you can ask it, right? I mean, it it it created this this wiki which is like a set of markdown files.

[00:59:31]Um, and and and you will see it here.

[00:59:33]So, the next tool that I will add here is I'm going to add something called if you want to play around with this over the weekend, that'll be fun. Um, go ahead and try out this one. I think this is super super cool tool. Um, it's by Toby. Toby is the um I think Shopify CEO I believe. Um so he released super popular um project. So this one basically it has like a lot of different algorithms that do semantic search. Like I said if you're working with um you know high density information like research papers like I think you you will need something with semantic search you know I think it's it's a given that you need something like this. Um but they also have full text search BM25 full text search keyword search right like LLM ranking everything is there.

[01:00:14]nice little like very compact um tool and I've been using this one like I don't need to pay for for something like this. This is free. It's open source.

[01:00:21]It's CLI by the way. It's very friendly with your agents. This stuff is amazing.

[01:00:25]Like go and try it. Just go and try it and and we'll we'll have like a agent lab around this. I'm working on this to to show you how you can use it. If you can wait a couple of weeks, we'll we'll surely have an example of this because this is the next tool I want to integrate in into our platform. Um this one is a little bit uh a little bit different but because it's super simple uh it's implemented very simple um you know um it' be it' be it'd be good to to integrate this one but play around with let me send you the link here um again this is alpha semantics stuff yeah there we go thank you for sharing cool let me see hab what you did here oh so this is on the on the cell okay got to see this actually worked Um yeah.

[01:01:09]So notice that notice habib I'll tell you like notice how the um the information that it's pulling uh this specific model haiku um is the laziest of all the models like it tends to want to produce very concise information.

[01:01:23]That's something interesting that I found about this model. Um if you know I think set for instance is much better at this. Oppus is much better at this. This this is why like I think this makes me makes me value more oppus like the latest like oppos 4.6 6 4.7 because those malls actually want to go deeper, right? Um it makes sense because I mean they're operating >> very expensive opus.

[01:01:44]>> Exactly. They're very expensive. Yeah.

[01:01:46]Mhm.

[01:01:46]>> Yeah.

[01:01:46]>> Um yeah, we're we're going to play around with different sandboxes. What one I I also want to get into the like Gemini malls. Gemini malls are really good for our use cases um as well that we have here and want to support and we'll see which other malls like maybe GPT malls as well, but that's more in the future. Anyway, so this is sort of the information that we have here.

[01:02:05]Again, you can see Haiku tends to like make things very concise, right? I mean, I can I can kind of nudge it and steer it and tell it to be more like more expressive, more details, right? Because I think that would matter here for this use case. But anyway, so that's not the point of it. I wanted to show you here the um the the the structure of the wiki. So structure of wiki. Oh, I I did I did ask, right? So let me just Okay, so here it is. So this is is is put in a wiki folder. Um, you know, this is assuming again the skill is assuming that you're going to have a bunch of skills. So, it's going to have one on self-improving agents and then it has a little configuration, right? What's the audience code blah blah blah anything about the wiki at a high level. Um, sources I think is great to keep track of on any knowledge base. Uh, that will help the agent a lot to find information to again the question about authoritative sources. You can actually kind of modify this to make the sources that you actually want the agent to use, right? as you continue to maintain the wiki, there is where you want to put that information. Okay, so you can kind of steer the model a bit. Um, and then here is like the actual information uh frameworks overview metac. So this is sort of like the structure of the wiki.

[01:03:12]It gives you a little picture of that.

[01:03:13]So yeah, there are markdown files stored, you know, in the sandbox and then um the agent produces the artifact now. So this is HTML plus the markdown, right? So it's both of them combined.

[01:03:23]And they may be wondering why why don't just inject it into the HTML. Well, here's the thing, and this is where I disagree with Tar, right? Um, for this particular use case. I It's very knowledge intensive. I need to use a knowledge base. I need to use like a database or something to to to keep that information to maintain it. Um, and I'm going to use Markdown because agents love Mark files. And so, the ALM knowledge bases stuff that Rapati was talking about makes complete sense for me. And this is what we we're we're showing you here. Um so what I'll do is I'll show you an example of like at the end of the day let's say we were very happy with this first version like it's has key information right we can keep extending on it right what you will do is you can probably like by the way this all of this stuff is really interesting like it's interesting that this one was highlighted this is one of the best papers that you can go and read on the on the subject of self-improving agents um there was a company I think in the UK that raised a significant amount of money on this stuff and this is one of their uh key key papers U so that's going this one as well I think it's by them as well but there's a lot of like interesting papers here that you can read on the topic but anyway so let's say like we we want like we want to understand we have an overview of this right what this is what's the timeline like right and so on this is really useful information we have like approaches as well uh metacognitive learning uh let's say we are happy we have in a good state so the next question I would ask is how do we maintain this right so in the wiki builder something I haven't covered yet is you have these like uh specific prompts. Let me share with you here and this is something you can try on your own. Um so you have like this maintenance stuff and you have this linting that you can do. So as your knowledge base grows, right, it's going to be really important to maintain this knowledge base because what you will see with language models and agents is that as they compile this information eventually because it becomes so complex the structure of the wiki um that it starts to make mistake about where to put stuff. So it does pay off when you think about the structure of the wiki initially um and and putting all of these different informations like config and so on. So you just don't want to put like like a set of files on a folder and so on because the agent would would not know what to do with it, right? So so this is what I was saying like we we we need to think deeper about standards and how we're going to build this stuff because it's not clear to me that you know when you scale this that the agent will be able to really take advantage of it, right? Um, so yeah, so the maintenance part is going to be an important one. Now, when I do querying, for instance, when I query the wiki, if the artifact was quering, all of that would be stored here. There's a lot of really useful information that's going on here. How to utilize the wiki, right?

[01:06:01]Maybe examples of prompts that you can try from the wiki. There's going to be derived information as well. So, as you put information together, there's going to be more der and there's good what you will notice is that there's layers of knowledge that you're you're building out here, right? And the higher it goes, the more useful that information is ideally, right? But again, we need to think about the structure of the wiki initially. So, so yes, um I have a really cool example that I think would be really fit for the the scale, which is um like take companies for instance, any companies that you want to track, uh research companies, you know, whatever whatever type of companies you track um and ask to create like a wiki of like the the the people, right? The products, the organization, the structure, the timelines, the releases, anything, anything about that. And what you realize is that this information like the agent is able to use it in really interesting ways because again it has this like structured information.

[01:06:58]Structured information is so powerful for for for language agent for you know language powered agents um because they tend to know how to use that information better. It's it's almost like you're it's almost like you're encouraging the agent um when you give it this you know structured wiki to look at the information in a certain way. Um to me again lots of research going on in this space but to me this is where where where where it it matters right the structure of it is is the thing that's going to be really powerful. Yeah Habib you had a question.

[01:07:28]>> Yeah does the update of the knowledge base is selfmated.

[01:07:32]>> All right so so let's get to that question. Right. So this as it stands is not automated. Okay. Um so if you guys want to try out what an automation automated wiki looks like you'll have to reach out to me. So, where did you reach out to me for this? So, I'm gonna I'm gonna go here and and show you in the community here. AI builder. Okay. So, in the builder, I'm going to um I'm going to post it right away. I think I can post it right away. Um I'm going to show you what it is right now, but I'm before I forget because I'm sure I'm going to forget this. Um if you want to try out the automation feature right away, let me know because I'll I'll I'll probably I'll have to enable you to try it out because this one this one does get extremely expensive really quickly. So, I want to like use some beta, you know, if you want to test it, some beta users here, um, you know, to privately test this stuff. If you're interested in that, just let me know.

[01:08:22]And if you have a good use case, I I'll enable it for you. Um, so just put your name here, a name here. If you want to test the automation feature, and I'll show you right now in a minute what it is. Um, I know I'm going to forget this stuff. Put your name here if you want to. Yeah. So, just put your name there.

[01:08:40]I'll find you in our database. I'll find you and then I will enable it for you if you want to try it out. Yeah. So, just put your name here, reply and the reply section here and I'll find you. All right. So, let me um I I'll create a Discord group and everything, right?

[01:08:52]I'll talk about that later. Um but let me just go back here and let me just go here. So, this is a little wiki that we were building and we're going to finish it with the automation part. So, yes, as I was saying right like this I would like it's a topic and it's highly evolving like it's evolving all the time. The knowledge base will evolve. A knowledge base by by the way needs maintenance and all of that stuff. We need to figure that out, right? But at least the automation part is going to be important part of maintaining it. So what I can do here is I can tell it something like this. So um I can tell it can you set an automation right to um update the wiki. Again let's assume that we did the work. We're happy with the structure of it. um let's make that assumption otherwise we're going to spend a lot of time actually trying to improve the structure and so on. Once we're happy with it let's just assume that um it's time to do some automation on top of it. So can you set automation to update the wiki um with new papers uh approaches benchmarks whatever that is right and benchmark okay um and use sources like archive for the question about how do you control the sources here is where your domain expertise comes into play I'll say hacker news hacker news always has like the latest news on this so maybe some um you know lots of companies around self-improving agents there's so many of them all over the place. Um Um and maybe like the Techrunch, something like that. Okay. Um All right. I I don't read Techrunch that much. X I do read a lot. Um so I don't have an X integration here. I heard they're using they have released some nice X uh X APIs, but yeah, I mean that's where I read my all my information, right? Related to releases.

[01:10:34]But just as an example, right? And and what I can do here. Okay. So set I want this automation to be every you can set automations for every day or weekly. So every day um at around I think most of the news happens in the morning. I'll say like 7 uh around I'll say 9 9 a.m.

[01:10:53]But you know I I can configure this however I want to be every day at around again. You're not pressing buttons nothing. You're just telling the agent the agent has this MCP tools. It's tools it's going to use. It's just going to build up. I mean the sandbox already does that for you, right? So it's quite safe to do it here anyway. So um so let me just uh Okay. Okay. So if there's anything, it's going to ask me if if there's anything that I need to clarify.

[01:11:15]But yeah, I mean every day uh at 9:00 a.m. it's going to look at the sources.

[01:11:19]It's going to use the web search functionality. It's just going to go around the internet and it's just going to maintain itself, right? Like I said, like okay, so let's say you trust what the agent is indexing here. you trust and you're happy with like okay you know the frameworks look really I mean at least it did a really good job at this stuff because it has the web search like I think all this information is good but notice it doesn't have anything in 2026 right that's right away I saw it no 2026 papers like that needs to be improved um and so on right so you have to verify it like you just can't build a wiki blind like I think you need that part um so that's again that's where your research capabilities your your domain expertise is going to count you can't just rely on an agent to put something like this together for you and expect to be of high quality source of high quality.

[01:12:03]uh in the last uh building not using uh wiki >> sorry what's the question >> in the last uh prompt you are not using the wiki builder >> um yeah oh okay so here don't hear you mean um yeah I mean it's going to it's going to know how to use it because what the automation will do is it's just going to it's just going to um load the environment and it's just going to take a checkpoint and it's just going to run the automation for you. It's quite simple, by the way. Again, lots of details of about how this works behind the scenes. Um, but yeah, it's just going to do that for you at 9:00 a.m. I think. Uh, but notice like it showed you something here. Um, and you might see some bugs here by the way because I I mean this stuff is really complex. Okay, the way how this works behind the scenes is super complex. I mean, the idea is simple like you know, you know, scheduling things like Nan, right? You can you can schedule with cloud code as well. it has like this loop functionality or something like that. Um and you can do automations with codex.

[01:13:03]That's kind of the idea here. Um a simple implementation, but I mean behind the scenes there's a lot of stuff going on. Okay. Um so so this one is already if I look at the configuration, how does this work? Well, it's a prompt and it's an agent doing the stuff. And this is where I think this unlocks so much stuff. And this is where I was able to do some of the use cases that I share with you here. Um like this for instance like it has all the context. It has the wiki. It has examples of the things that I was storing in that wiki. It has you know all the all the historical information right um it takes all the new information and then it it has again it has a structure to guide itself right of the wiki and then it's just self-improving itself like it's just self-improving self-improving and it's just incorporating and you know the more it does it the more information it has like the higher the quality of the information that's coming in. That to me is where like that's a a mind-blowing experience from LLMs because LLMs we we know it right like every time you try something it doesn't memorize something or something like that or hallucinates or it's making updates or something like that. But the wiki and the structure of the wiki like I said is kind of encouraging the agent to look deeper at the structure and to follow what is stored in in in in that wiki structure.

[01:14:16]Um that that to me is just incredible.

[01:14:19]like I I think this is where um you know wikis are extremely powerful LLM wikis in particular. Um but anyway so it's is a you know you can see it here it's uh this is the the prompt. So it went from my prompt and then it made a detailed prompt because it understands now okay I have a wiki I have something that you asked me to do and then it's going to say okay search for agents and it's just going to use the tools the web search tools and and it's going to maintain this it's going to happen at every every you know every day at 9:00 and again it's all a prompt we don't need to in any way you do this is like you have to take a node you have to you know connect this connect that connect no this stuff is just a prompt it needs to be really simple as far as like as as as far as I'm concerned like my opinion about this it does not need to sophisticated. It needs to be an agent doing it. Um I think there's so much complexity that people like implement this stuff with.

[01:15:08]It doesn't need to be. My best experience of this has been just do it as a prompt. That's it. And an agent and that's it. As a sandbox that solves your problem. Okay. So, um yeah, look at it.

[01:15:18]It has the wiki sources, right, with the new entries. It's nice because it understood that context. And then it has update the log maintenance and anyway.

[01:15:24]So, I'm going to say yes. I'm going to that's saved, right? Um, and what I can do now is I can just run it. So if you want to like see how it works, you can just run it right away. Um, let me see here. Uh, so the source is 2035. I I wish it would pick 2026. I think it's haikum is probably going to be a rough one to to get there, but I maybe I can kind of steer it a little bit to get there. Right. So let me just run this and see how it um for you to actually see it in action. Um, yeah. So you can just run the agent once. This is something you can do anything. Just run it, test it, and see what it produces out there. um you know how it updates things. So let me just go back here. So I expect to see something maybe maybe new papers things like that right they will update and then I will ask the agent what what actually changed this one take a couple of minutes again I'm sorry for taking more time but uh maybe in the next five minutes this is be completed I'm going to end the session and we'll share the recording later u any other questions from the group whoever didn't get a chance to ask please feel free to ask a question >> yeah I have a question if I have the PDF of a paper from archive or API to archive paper. Can I get a interactive uh artifact?

[01:16:39]>> Oh, like um you mean you mean for um if you want the full you're referring to the full text, right? Um >> the full text of the paper. So we can we can we can do that. We can package it. I have something that I built locally already. Um I think if there's a lot of interest rate, it it really depends on the group, right? if there's a lot of interest for it, I can probably build um I can probably build a prototype that I can share with the group. Okay. So like it can be something that I can share that's uh for everyone accessible. So we'll work on a few of them like lesson generators things like that and we'll share with they'll be sharable, right?

[01:17:14]So you can see them in in the builder landing page. You will see all of them and you can interact with them and you can go just kind of expand on it. Okay.

[01:17:20]So I'll build something like that if you if you guys are interested in it. Yeah.

[01:17:26]Any other questions from the group?

[01:17:28]anything you're curious about as we get the final results here?

[01:17:32]>> Yeah, I I have a question too. I don't know.

[01:17:36]>> Go for it.

[01:17:38]>> Um, can you hear me?

[01:17:40]>> Yes, I can hear you.

[01:17:41]>> Yeah, I you mentioned some NCP tools like MCP tools to like the the agent already has MCP tools >> to make the automation work. Do you have some specific ones that you can share or do you make them yourselves?

[01:18:00]>> Yeah. Um um I so I have I have two different types of MCP tools that I'm using here. Um so the first one is um the web search. So I use XA MCP. So XA XA is web search. Um so I use their MCP I find their MCP tools really good. Um so I'm using their one. Right. So that's the first and this is not something I built. This is just like I just connect my sandbox to the XMCP. The other one is um it's so these these automation the automation there's a set of MCP tools in the automation in the sandbox that is doing the automation stuff. So you can like create the automation. Um basically it's just a script like I think it uses Python as a script that it uses. Um and it uses a fast MCP and then it creates these uh it uses cron right behind the scenes and it creates this automate this this this automation. So it creates it can edit it can delete an automation it can configure the automation uh pause automation resume it so you can do all the things that you would do around the automation. Um I don't think that's the the um like I don't think that's the best interface for the automation. My opinion is that everything is going to go to the agent anyway. So, might as well I build it as an MCP. Uh but if there's interest to that, I I can probably do something. I can probably share more more details about that how that works. Uh but those are the only two that I'm using here. Um we are going to be using uh a few other MCP tools.

[01:19:26]So, another one that I've been um that I built is and this is in the hands of LOBS stuff that I that I shared earlier is um image generation, right? So, image generation, it's using um an MCP server that I built. this MCB server is connecting to to nano banana. So the latest nano banana model which allows you to generate images. So that's another one that we have built. Um but those are the only three that we have kind of built for our platform. Um and again if there's any other tools that like that that you think would make sense on our platform um I'm happy to like maybe support it uh and and work um you know work on it to again to allow it. But you know I think the the only thing with the MCP tools is that um um if it's using you know like let's say you wanted to connect your inbox or something like that. So we have I have another like tool that I've built like an artifact that I built. I don't read emails anymore. Like I don't like I use this uh Google MCP tools and basically I just connect, you know, to my Google accounts and my agent connects to that and then I just get like a like a little um like an artifact and gives me basically it just filters out all the spam stuff and it gives me all the important information. I interact with the artifact. The artifact I reply I use the agent to reply sometimes. Um and like in 15 minutes I'm done with my emails. Like I don't read emails anymore. that is the kind of thing that the stuff is really unlocking already.

[01:20:44]Um and again MCP is is kind of you know I trust MCP because it has all the authentication all the stuff in place.

[01:20:50]So it's really important technology I think. Um we don't talk about it a lot here but but I think if there's interest in that I'm sure we could do something around that and we could share more.

[01:20:59]>> Yeah. Um >> thank you.

[01:21:01]>> Okay. So automation field. Sure. Sure.

[01:21:03]You're welcome.

[01:21:06]>> Okay. To mention Karpasi is now uh uh with with Antropic is working with with Antropic is living.

[01:21:16]>> Yeah. Yeah. Yeah. Yeah. Yeah. It's very interesting. Um very interesting. Um yeah, development. All right. So, this didn't work for some reason. I'm not sure what happened here. Uh this should be working. I probably have to look at the logs. What's going on here? Um let me try again. So, let's see if I can reload this.

[01:21:35]Yeah. Yeah. I mean, it has a few bugs here. Keep in mind this is like a better like better feature. I I don't uh it's not perfect. It's going to have issues like that. If you see issues like this, please report them in the community. Um I'm you know we're we're constantly trying to improve this stuff. This is very new early days. We just want to make it very useful for you all and if it unlocks use cases for you, wonderful.

[01:21:57]But um but yeah, anyway, so it definitely should work. Um I've I've used it a lot already. And um you know I I'll probably do like a separate recording of this showing you how it works. Um but for some reason it didn't load here. Let me see if I can fix it.

[01:22:12]Uh but this will take a bit more time. U yeah I just ask just ask question.

[01:22:38]You're welcome. You're welcome.

[01:22:43]Yeah, I mean the wiki um so Kiki asked question about how much of your daily weekly budget is spent on creating um updating updating the wiki. So I have all of these wickets that I've built, right? As I was saying, but you know, I've been using um I've been using like I use my my my subscriptions, right?

[01:23:02]This is the powerful thing about having a like a Mac subscription on on Tropic or using your codec subscription as well. If you have Hermes open agent, open cloud, whatever you're using, you can definitely use those to to build a wiki. I mean, at the end of the day, it's just an agent, a coding agent. As long as it can do code, um you know, it's good at code, it'll be able to build it for you. But I'm using my subscription plan so I don't actually feel it that much. And I have these automations that run like every day.

[01:23:27]Like some of them run like every four hours. Uh it really depends on the topic. Like emails for instance I I would have that running like you know my inbox zero stuff is is running like every couple of hours. Um yeah automation really depends on the use case. But um but um yeah um yeah I using my subscriptions. I don't I don't pay API credit for that. I just use my subscriptions for that. Yeah.

[01:23:50]Yeah. So, let me just go back here. See what's going on with this one. Um, it took me out there.

[01:24:02]Yeah, there's definitely a few bugs I need to go and and figure out. Um, but most of it should be working. So, let me see. I had this simple wiki here. I can probably try it here.

[01:24:14]Um this is much simpler. I think the the the issue is the um I think I really understand the there's some limitations on the um on the automation thing. The thing with the automation is that it basically going to take a checkpoint of your sandbox and it's just going to run, right? It's going to try to run that. So it can be depending on the task itself.

[01:24:38]I think the task was a little bit more intensive. So it can use like 50 turns, 100 turns, right? So if it's going to use 50 100 turns, I think I have a limit on that already. Um, so again, if if your use case requires that, please talk to me and I'll I'll enable that so you can kind of experiment with that. Uh, but automation is not it's not publicly available yet. I just wanted to show you how it actually works. So let's set automation um every day at nine. I mean, this is the stuff that I I really wanted to show because this is the stuff that makes it super super interesting because at this point you just have something that's just updating itself, right? and and and and you get informed on whatever topic and and again you filter out everything else, right? We need we need more solutions like this. You want to filter out all the noise. There's so much noise going on in the space like it's really hard it's really hard to to keep up with things like that way. Um okay. Oh yeah. And in talking about that, like I do have one another another wiki I've built like it's on X because I don't tend to consume X the same way I used before. And I'm using the I'm using their developer APIs and I have this like bookmark functionality. Sometimes I would go on X like five minutes and I bookmark um helps me build my newsletter. So I have this bookmark functionality and then I have this agent that has this automation and it goes like every morning and it collects my bookmarks and then it just like you know helps me with all the topics that I want to talk about in my newsletter. So there's a lot of automation like that happening already. And what this has done is it it basically I don't open apps anymore like I don't open the browser like I'm on my ID working all day. I don't care about the like you know if if if X is has too much information I don't want and there's so much noise I filter out already right and and my expertise helped me again build the automation that I want this is why I said like that's not something you could you could just replace easily like your domain expertise but um just just another example like nine do let's see if we can get it done um if we can't then I'll just probably do another we'll probably do another we'll do a couple of them every week I'll try to do one of these as we introduce more functionalities. So stay up to date on that. Um visit the community. Um I'll I'll announce it. You probably get emails if you are already have an account with us. Um but we'll do a few of these based on the questions that we got today. Um some of the interests, right? Like how do we make it more pay attention to authoritative tools, the MCP tools, I think was an interesting one. Um the the the rag stuff, how do you combine rag rag? There's so much stuff like we we we can talk about this stuff in depth, but the one that I do want to cover is the one on human verification that domain expertise stuff because every application I think that I've worked on, there's always some human input that's going to happen and that usually gets you like sometimes you get 90% from the agent, but the human is going to take you to the 9 to5 which is what you need to really deploy things into production. That has been my experience since the beginning of language models working with companies.

[01:27:33]it has always been the case and and and and it continues to be the case here.

[01:27:36]So, this is something I want to kind of continue to push a little bit more. Um, so you'll see more content around that.

[01:27:41]Anyway, so we have an automation, I do believe. Yes, we have it. Uh, let me see this. Okay, so this is a shorter one.

[01:27:46]This is much better. Um, I'm going to try to see if I can get this to work here. If it doesn't, I need to go and find out what's going on.

[01:27:54]Yeah, but I think it's a it's a it's a turn thing because I have a limitation on the turn uh on the number of turns.

[01:28:01]Okay. Um, Hermes agent. Um, Hermes agent. Yeah. Um, so I was like, okay, so before I made the decision on which of the sandboxes to use, because how do I was going to build my sandbox, right? So I actually run a few tests on my use cases, which is what I'm showing you today, like the wiki, the skills, and all that stuff that I have that I work on, uh, the research use cases, uh, the marketing ones, uh, being able to work better with MCP tools and so on.

[01:28:29]Um, so I've tried a few of them. I think for me Hermes is a little bit too complex in the way it works. Um there's like this self-improving stuff in it. Um you know like it's it's like open cloud, right? It has this kind of soul empty.

[01:28:41]It has this thing that um and and and to some to some degree it does drifts. I notice a lot like it does doesn't matter which model you use. It just has all of these like this structure, this information, this memory stuff, this all of this stuff that is packed and it's way too much. So for my use cases um I need to be I need to have a tool that you know I can modify really well and that doesn't have too much information.

[01:29:05]Cloud NMD file is good enough with a set of other files and a wiki and that's good enough for my use case. Everything else is just kind of noise to the agent and it doesn't perform as well. So from a harness point of view I do prefer to use cloud code and codex. I think codeex is where I'm leaning at more these days in terms of like code code like you know software engineering code bases and so on and when it comes to like marketing product stuff building prototypes uh research is going to be more leaning towards cloud code um and using cloud agent SDK as well um and and then for the rest of stuff just more experimental like longunning agents I use cursor and to some extent open code and I've tried Hermes I haven't had good results with it um but yeah I'm I'm going to I'm going to explore a little bit more on it. I think I haven't went in depth in it and really explored the ecosystem enough because I mean I I spend most of my time on Codex and cloud. That's where I kind of um you know have been focused on the past few months. Um so unless there's like something really really drastically different from Hermes, I'm not too sure and convinced about that yet. Um yeah, again I I've heard good things about the memory stuff and so on, but um you know I think for memory is like the simpler it is the better. Um that's my opinion about it. I I don't need too many sophisticated files uh on on on on my system to to get the work done. All right. So I think we had we got something here. Let me see what we got here. Yeah, this time it updated.

[01:30:31]This time it works. So I I think it has to do with the turn stuff. So I'll have to kind of increase the the the the rate limits there. I have to adjust those rate limits because I think um that was the issue with the with the previous one. But anyway, so this one you can see it worked. Um let me see here. Uh I'll help update. I found a substantial new research. So let me just go ahead and check. I think let me just refresh this if it if there's anything new here. Uh working memory which strategies membrane. Okay.

[01:31:06]Okay, I'm just trying to find where where did it it it doesn't look like it had it updated it though. Uh maybe I'm missing that completely.

[01:31:18]See? Oh, okay. So anyways, so um so the reason why this one didn't work is because is because like this one didn't use the wiki, right? So this is the thing that I missed completely. So this wiki that was built was just an example using the front end design which is basically just an HTML file with some text. That's it. It doesn't it doesn't have like it doesn't depend on the on on the wiki stuff, right? So the one that I built which didn't work which is this one here which I wanted it to work. This one does use a wiki. So the whole automation stuff that I did um just assume that there was a wiki but there wasn't one for that initial one. But anyway, so the automation stuff does work. It did pull information using the web search. Um, but anyway, so now now it's like showing up here, I guess.

[01:32:00]Yeah, there we go. I mean, again, you see like it loaded now. I just needed to reload it, I guess. Uh, but there it is.

[01:32:08]Um, again, it's not using a wiki because there's no wiki here. Um, so that so but that was my fault with the with the with the I guess through the prompt. It's just an HTML artifact. That's it. But anyways, you can see the information being pulled. It updated, right? So, it it actually worked. It's just that I didn't load it. Um anyway, so I'll leave it at that. We'll do more sessions on this, if there's interest. Um again, just let let us know if you have any feedback. Um this is all very experimental. I hope you can use it.

[01:32:37]Give us some feedback on it. If you have any questions that do come up, if you have a use case, let me know and we'll try to see how we can support it. But I'll encourage you guys to go and try the wiki, go and try the lesson builder if you want to learn about topics and we'll integrate that into our courses which means you'll be able to generate quizzes and stuff on top of it and we'll be able to use the coding agent to help you along. Okay, so thanks so much everyone. Thank you for your time. I'll see you all in the next one. Probably next week we'll do another one and the week after as well. Um we'll do more live sessions. We'll do more build sessions as well. So for those of you that have the pro account um we'll do a few of these and we'll build together.

[01:33:14]Okay, so we're going to be a few sessions like that. We call them the build sessions. They're quite popular, but we're going to do a few of these as well. So, that's kind of the plan. Thank you so much, everyone. Have a great day.

[01:33:23]I'll share the recording later. Um, have a great weekend ahead. Thank you. See you all soon.