How to Structure to Data with AI: Interview With Kirk Marple of Graphlit

not just another GPT rapper. And in today's episode, my guest Kirk Marple walks us through how he is building AI and content platforms for other developers to build systems on top of. I thought this was super interesting to hear about not only his career journey and his experience, but how he is looking into building these types of systems so that other developers have other tools to build on top of and build even more complex systems leveraging AI. This landscape is changing dramatically all of the time. So, I always find that it's incredibly interesting to learn from these other individuals who are pioneering and pushing forward. So, sit back, enjoy, and I will see you next time. Kirk, if you don't mind kicking us off, then would you like to give us a little bit of background for for where you got started? And like I said, you can kind of jump in as early as you want or however you'd like to frame that up. >> Yeah, sure. Um, let's see. Kirk Marple, I'm a founder of a company called Graphlet. Um, been in tech for 25 30 years now. um started off as a developer um really I mean from early age I mean just I graduated college early and went for CS and then ended up in grad school in Vancouver and ended up coming down to to uh Microsoft um so kind of Microsoft was kind of postgrad school world and then had startups for the last 20 years or so um always been around media technology video audio stuff like that and then it's kind of blossomed into unstructured data and um and now AI is really another way to consume all that. But I've done everything from I mean audio and video transcoding to um I mean drone imagery and I mean just everything all around the board. So >> that's super cool. And I guess I have a couple questions and so going from from Microsoft a big tech company and then kind of shifting gears like completely into startups for like a large majority of your career. Um did did you find that you were like in big tech and you were like h not for me and just like really wanted to get into startups or was it just something that kind of like happened a little bit more naturally? >> It's interesting. I mean like I so after college I moved to Washington DC and and initially it was like very government heavy like I was working for somebody they work for the Marine Corps and I mean it was cool stuff though. It was like um mapping software and things like that and learned a lot. But it I mean I I ended up just falling into a couple companies that were really interesting and you you'd probably call them startups now. I mean they're some of them had been around for a couple years but they weren't big. I mean they were less than 100 people and >> I think I think the first three companies I was at were all that size and then I went to grad school. So honestly I just I had had a bunch of maybe four years of of kind of startup size company experience. Um but I loved Microsoft. I I was on a great team. Um I was part of uh basically what was the Microsoft network to kind of date myself back to that those days. And then um was in Microsoft research in a really cool um we were doing 3D virtual worlds back in the in the day which was like a super awesome experience. And >> and I was checking your LinkedIn and so the date for that that would have been that would have been a very early time to be doing like that. That's super cool. I mean, it was it was before the Yeah, it was like '96 um around that era. Like, we had a paper out and if you think back, I mean, what's that? Gosh, it's 30 years ago now. But it's crazy. I mean, that was like I mean, multi-user um like domain or dungeons and like um game multi- um player gaming and all that was like the old I mean, originally. And it's I've seen things come out now that I thought of back then and that are now even just kind of coming out. So, um, but to your point, it's I I love my time there. Um, at the end, I kind of was getting, um, I mean, I think I was wanting to talk to the customer more, and I think I was I was feeling a little bit separate of just like, what are we doing? Like, who are we doing this for? >> And I was like, okay, I had a feeling I just wanted to go to a smaller company or a startup. um and ended up with I mean I jumped I jumped out um right about in 2000 and then ended up starting this company with a guy that I met and it was a journey. I mean it was a lot of ups and downs and I mean almost ran well I basically did run out of money a couple times and then but I mean I have all the stories of like bootstrapping and going to have to contract for a while and coming back and I mean then getting a billion dollar contract and it's just like all the ups and downs and ins and outs. So u but it was I mean I wouldn't I wouldn't trade it. >> Before we move on, this is just a quick reminder that I do have a course on C refactoring available on dome train. Refactoring is one of the most critical skills that you can learn as a software engineer. And this helps you continue to build upon applications that already exist, making sure that they can scale and have extensibility. I walk you through a bunch of various techniques and give you some examples that we walk through together to see how we can apply these techniques to refactor the code. Check out the pin comment and the links in the description to get this course. Now, back to the video. That's so cool. And yeah, like you know, tons of like really amazing experience that like basically must come from that, right? like when you're kind of forced into doing that. But there's one more thing I wanted to ask about too. So at Microsoft you were a manager. Um and >> so I I'm just curious from your perspective. So you would have started off as a developer, moved into management. Uh so the two questions I have about this are like what did you find like that transition was like going from developer into a manager role? And then I know this is going to be a little bit of a weird question to ask the other way cuz when I look at your other work experiences, a lot of like working in startups being like architect and this kind of this sort of role. So moving away from management back into maybe more of like an I want to call it an IC role, but at the same time I realize if you're running companies and stuff that's not really necessarily an IC. So I'm just curious that transition. >> Yeah. Yeah. It's I mean it's a really interesting thing. Like I I'm definitely a builder. like I enjoy the development, I enjoy building, but that leans really blends into product as well. And so um I mean I started out developer, senior software engineer, like whatever you call it back then and then I became a dev lead at Microsoft, then I became a dev manager at Microsoft and then after I left I was a director of engineering for a little while and then jumped into the startup. But once you're in the startup, you're just wearing all the hats. And so that's for me and and it's like I'm almost unemployable at like a normal job now because it's like I love it all. I mean I love I love doing the product. I love work with the designers. I love the AR software architecture. I mean I do all the a lot of DevOps too and I'm more of I'm more of a backend dev I'd call myself and um but I built UI and like Zen I've actually been working on the front end stuff with this new product. Um but it's I mean I think the thing is where when I have been CEO or when I've been CTO I what I would do is kind of look at what my skill set is which is usually data modeling architecture systems distributed systems and lean in there but then hire people to do the other parts like the UI part or the design part or things like that. But I'm more of a like I think the the builder side of CTOish like that kind of world where I know some CTOs that are like super into the um like building their employees and building their teams and like the people side of that. Um and while I've done it, I don't necessarily enjoy it as much and it's not my natural. Um, and so I'm I I fit I think better in the 0ero to one stage where I just love talking to everybody at the same time. Have a very flat org just like it's like there's no there's no like senior junior whatever. It's just like we all work together and so I love I love that kind of environment. >> That's awesome. Um my like we haven't chatted about this but my work experience before Microsoft was at a a startup that had grown into a like a 250 plus person company and >> I remember going through that period and like you said wearing all the hats. I remember being like I don't care what my I was a manager but I'm like I don't care what my job title is. I we were doing everything together all the time and then I came into Microsoft and I realized like everything is about your level and your title because if you want to be you know promoted if you wanted to switch companies like you need to be able to say I'm at this level so like >> it's a lateral move or I'm going up like >> it's not something I ever had to consider but um it can totally relate to what you're saying for being like in that startup kind of phase and just being like got to wear all the hats and it's uh It's I find it's really exciting. Um I thought something else you said that was super interesting was >> like you kind of have to you have to know your strengths and what you want to be focused on. So from from having to run companies >> you're in this position where you're like okay this is what I want to do. This is what I'm good at and these other things need to happen. >> So you have to figure out like I'm assuming you have to figure out for yourself how much of that you're going to do before you're like okay this is someone else has to come on for this. And I mean that's I could talk all about the sales and marketing side because that's really the part that I I mean coming up as a dad like I had no knowledge of like what I mean sales and marketing and I took no business classes and all of that I learned on the fly and just from watching other good VP of sales or sales directors and things like that but but it's also an area I made a ton of mistakes in. And so you try and delegate that because I'm like oh it's not my specialty. I'll hire people to help with that. And then you end up that they're not they don't get it the same way or they're not as dedicated or they're just the wrong person. I mean, they may be great for some other company, they're not great here. And I mean, and I think really the marketing stuff I realized there's more of a process to it. And it took having a good marketing person that we had hired part-time for me really to learn what that means and like about the processes and stuff like that. And a lot of it's common sense, but you kind of have to see somebody go through the the motions. And then and so now it's like I end up doing a lot of it because we're we're a developers tool um developer platform company and like it's really more developer marketing and community building and social media and it's not classic like sales at at this point. But my previous startup which I had for 10 years, I had a a business partner who was the sales guy. like he was a great a great relationshiporiented like he'd take him out to dinner and like and I could just focus I was like CTO but like fully just dev customer support like that's all I've been um and but it was it was very like he was selling to ESPN and Fox and NBC where it was very much who you know relationship kind of stuff. Um, but the only downside then is when I started this company, like I didn't, it wasn't like that at all. Like it was very organic. And even now it's like I mean I feel that we're still blocked on how big we could be by just clicking on that like that sales and marketing more. >> Very interesting. And I like I have nowhere near as much experience in terms of startups and you know running companies and things like that. But even for the stuff I build on the side, um, you know, I've had one thing before that failed after a couple years, like this is again just part-time on the side building stuff and really realizing like this is not going to work if all that we do is try to write code. Like it it doesn't just sell itself like it does. You don't put something on the internet and you like, you know, look overnight and you're like, did we get a million downloads? Like doesn't happen. Um, and we're realizing that we have to be doing more and more of like the sales and the marketing because it doesn't just happen on its own. So, >> and it's so true and I mean even with product hunt and all these different things and I mean and it's I mean I almost get more frustrated by how slow the marketing growth is compared to the I mean the product we have under control like I built product for so long. I think if we have a vision of what we're building, like we can get there and and even now with AI coding, um it's 10x faster than it was. Um but there's also now I'm learning there's downsides to it as well. But but there's I mean I think the marketing is something that if you're in Y Combinator or if your VCs promote you really well or different things, it's almost like a class system a little bit where it's there are some ways to sort of get more visibility. Um but if you don't have that, you really have to work at it. And I I see I don't think it's um it's I mean it's not based on the product. It's really based on a lot of times it's I mean your ability to market yourself in a brand. And so I've been trying to lean in more on on Twitter and and previously more on Reddit. Um, I don't do much with LinkedIn anymore, but it's I mean that's where I've been trying to really get in front of people and just I mean I wish I had 50,000 followers like because I think it would it would make it so much easier, but it's a grind to just get get that going and and um because I think I mean I see that what people are talking about and I think what we're doing is super relevant. We're right on the edge of I mean MCP and Gen AI and data unstructured data like what we're doing I think is pretty cool but how do you how do you get how do you even get people to know you exist is is the hard part >> right and then to your point like you know if you had such an audience like that also doesn't just happen overnight so it's like if you wanted that you're like okay I got to go spend that time investing into building that yeah that's uh it's a pain >> now and and I think it's I mean that's what I've I've tried tried to lean in more to the sort of building in public kind of just talk a lot more about it and I mean I like I love being on podcast and just kind of be able to chat about this stuff but it's it's the daily and I do get people coming back to be like hey I heard about your product I I listen to this podcast which is great >> but you really it's it's it's a job like you really have to keep up with it and and I mean I can kind of see from our signed up numbers how much I'm leaning into marketing like this last month I've really been heads down building this new product And it's like I mean we we're getting consistent signups but it's not the same when I was like pounding like right like social media every day. So >> yeah it's and then it makes you kind of go like okay well how much time do I have to dedicate to that because you know that needs to happen for growth but then >> how do we develop the product? Yeah. So very tricky balance but maybe we can talk a little bit more about what you guys are building where it's at. Um obviously as much detail as you'd like to get into. I don't know if there's like secret stuff you don't want to share, but um like what's going on and uh and maybe you could even start with like >> how did how did you come up with like like trying to like work in this space and like this is an area to invest into. I mean that's what I what I was saying like I've always kind of been a data person and I had this idea of really kind of leveraging knowledge that we have around us and so when I had been working with the media companies they it was really about the eyeballs like I mean we were just putting putting video on Netflix or putting it on YouTube or things like that um which had its own set of problems but it wasn't about what's in the in the video in the data and I started to think about podcasts especially and video recordings and all that and and got really interested in in knowledge graphs again. And I had done some of that in college, but it kind of came back around. And I was like, "Wow, this is really interesting how all this connects up and started exploring how to kind of create the knowledge graphs from the media through like audio transcription and computer vision and and I'd started like I was at at General Motors for a bit and then I was at this the drone company Casper for a bit kind of building platforms like this to analyze media and I really just thought like I mean there's there wasn't a snowflake for unstructured data like there wasn't just a common platform this is like five or six years ago that everybody was using for unstructured data and not even just for like you wouldn't say like data bricks or something like that for ML stuff but just just managing like and I I used to call it like photophoto for industry like like what's the what's the tool that you could use in a business sense or in in that that you could leverage AI but be as easy to use as like ano or something like that um and then I had this idea originally what was this product called Zen that is okay how do you get the data and organize it and repurpose it. So, it's funny because now like what you've seen with Google with Notebook LM of, oh, you can create a podcast version of this PDF. Um, that was very similar to some of the ideas I had five or six years ago of like how can you take it in one form and manipulate it into a different form and kind of like here's my email. How can I get an audio summary in the morning of what happened overnight? And kind of that was where it all kind of bound together. And I actually started like mocking up this Zen app with a designer back like I mean before starting this company. This is like five or six years ago and kind of left it. I mean and then we ended up starting I mean ended up starting uh the company and took some of the ideas from there and we built an app that was um it ended up being more geospatial leaning. So it had like a map view and and it was very computer vision focused and it wasn't as document focused. Um, and it was really cool, but we couldn't, we had a hard time selling it because people were like, well, we're just, we're using um, SharePoint already. Like, why would we need this, too? Or we're using ArcGIS for geospatial. We don't want another pane of glass to have our have our people look at. Can you integrate it into that? And then even then, when we did integrate it, they were like, uh, I don't know. We already used up our budget on SharePoint. And so, it was like they they would say they had a problem, but then they didn't want to spend the money on it. is a classic thing and >> and we ended up going and and just going down a rabbit hole where we just couldn't get enough traction and ended up having to lay the team off and finally really hard pivot um from a an app company to a platform back to a platform company and since I had built the platform myself um ended up just having to like go heads down I even I mean got a separate job for a little while and kind of came back a few months later, I think it took about 6 months, rebuilding it as a pure development platform that people could like get API keys to and built out a new portal and launched that and so it was like 2 years into the company and then about 2 and 1/2 years relaunched it and that's what today we sell as graphlet. So people are building applications around it, building agents around it. Um, it's kind of rag as a service in a way where it's like you can ingest anything. You can integrate with LMS for like prompted results, but it's a lot more than that. It's um essentially media management integrated with AI. Um, and so then continuing the story now it's about 4 years later, what we realized, what I started to realize was like, okay, well, we had had this idea for the app originally. I had had the idea for Zen a year before that. could we use AI coding like cursor and and cloud um cloud code stuff to rebuild the app? Like how fast how long would it take? And we had already had sample apps in Nex.js and we had we have a really nice SDK. And so literally I just was like okay I'm just going to try it and see see how long like how long this might go. And it started to be go so fast in the first week or two where I was knocking out features that we had taken months to do >> in like a day, an hour. I mean, it was insane. And the good part about it is I had a design. I had an app we already had that I could like literally take screenshots of that I mean of of from old Loom videos and stuff like that. I'd be like, "Hey, make it like this." But what I ended up with was like a better version of our old app and now with AI and LMS fully integrated. Um, and that's what we're going to release hopefully hopefully this week. Um, and it's it's kind of there's like ano view kind of to it where you have your con what we call context view and you have your chat view and you can kind of you basically can ingest all your data um curate it, collate it like um create collections, create filters and then you can basically chat with that subset of data. So you can be like, "Okay, pull in my email, pull in my Slack." Um, now search for only data about Microsoft across all those data sources and now chat with that data. And but it also is an MCP client. So you can plug in like I was playing around with like linear is MCP and Century's MCP today and making sure that OOTH works and all that. Like that's a lot of the testing we're doing now. Um but the important part is now we can expose your data as MCP to other clients and that's I think really the magic here is now you can curate your own data set and have different views on your own data and then share it with claw desktop or share it with cursor or something like that. Um, and so it's this circular thing and I I mean that's a that I think nobody's really done that well yet. And so I'm hoping that is a big deal for the industry because it's I mean be able to curate your own data sets that you can expose with MCP while also taking advantage of MCP. Um, and so it's really a mix I would say of like chat GBT and notebook LM. Um, but really heavily on the content and context management part. Um because it's like I mean like I don't know I think we have like 15 different data sources you could pull from. We have like uh web crawling integrated podcast search. Um but also we've done nice things for like uh direct capture in the app. So you can actually capture like I could be I mean have this I could have a transcription running. Um, and I was even playing around with some of the like AI kind of um almost like a I was kind of calling it a guardian angel, but it's like it's it's listening to your meeting and suggesting questions and different things like that. >> Um, okay. >> Almost like what Cle and some of the other folks are trying to like have done, but it just fits really nicely within the kind of architecture of what we already have because you essentially have a knowledge graph on your data that we can retrieve from and suggest things. Um, right. So, it's really it's it's come full circle, but it's nice because it's built completely on our our platform. Um, and really just using like assembly AI for transcription or clerk for off, but it's also an example of what people can build with our platform too where we can say, oh, well, you want this little piece of it, you can go build an MCP server today. You can build a UI today. Um, and we may open source a slice of like somewhere kind of beneath the the full app at some point, but we also have open source um, sample apps that we already I mean show some of the stuff too. So >> that's super interesting. And I mean a whole lot to unpack there, but one thing that came to mind that I thought was really cool is that >> you have um, like essentially this platform, but you're also building on top of it like more product features on top. So you get to use your platform to build things. Um, in terms of that kind of feedback loop, how do you find that's like I'm assuming it's very helpful, but I'm curious to hear your perspective about how you use that to kind of drive platform features and that kind of stuff. >> It's really interesting. I mean, we we love getting feedback from customers. I mean, I think that's the part I love is somebody on Discord says, "Hey, can you do this?" And I'm like, "Well, we can do it like 80%. Give me till tomorrow morning or give me till tonight. I'll get it for you and like we'll build it that fast." And the good I mean I love being that responsive and and to customers and that part I think is fun of being a startup but I think with with Zen it was interesting like there's things that I like I thought were finished and worked perfectly but once I could build it into an app and do it the way I want it from a UX standpoint I was like oh wow this could work a little bit better or I mean or this one thing didn't really work the way that I thought it did. And it's crazy like as much as we have I was looking I mean we have almost 2,000 automated like integration tests and I mean you're running all these things but there's still until you really get it into an app and think through the UX um you just there's things you don't think of. So for me it was I I there was a lot of polish of just okay we can make this easier make this better. Um, and there was even a couple features that I had had on the road map for a while that I was able to finish because now I had an app to kind of like integrate and and work through that that whole dynamic. Um, especially around ooth and connectors and and that kind of thing. >> That's super cool. I I know that you had said earlier like from your perspective, you kind of consider yourself more of a you lean more towards the backend development. And I know with some of this you said more in the front end now, >> but uh I was thinking about this because for myself I'm very much the same way where like I love to go build systems and when it comes to the front end I'm like I understand user experience but if you want it to be pretty like I might not be your guy. Um but when I go building things sort of from the back and building a platform building out services and things like that in my mind I'm like this is probably rock sol. And then when you start coming from the other end, you're like here are all these gaps or opportunities to go improve. So it it sounds like you really do like from your perspective, you get really get to close that that feedback loop then by building on top of your platform. >> And I think it's there's a I I say I would I mean I think I'm decent. I mean I'm pretty good in the UX side of just having a good gut feel for that. I'm not a designer. I don't I mean I think I know what I like but I think one of the things I've had fun with in I've been I started using Klein um in VBS code and then I was using um been using cloud code a lot but be able to explain what you're what you mean in design sense and I've always had a hard time working with designers where I'm like I mean like I want it like this but it's like how do you put that into words and and like a lot of times I'll just like go into Microsoft paint and copy and paste some stuff and give just something to talk about. And I think it's been interesting to see how we can do that with AI of I I screenshot it and literally we'll like move things around and then give it back to the AI to be know okay I mean kind of more like this. And that's been an interesting learning experience is like how do you replace that human iteration loop? Um and I think there's so much green field that could be done better in that space. But um but also it's interesting because the the models kind of have a design aesthetic where it built some really nice stuff that I couldn't have told it how to build but you know it's like oneshotting something from some some GitHub repo somewhere but it actually made something pretty nice and and then but other parts of it I'm like taking my input and being like okay no I I have an aesthetic where there's like a sidebar and I want this panel and I want the sliding panel like and so I I was trying I was trying to find that right balance. Um but yeah >> and then yeah so you're super interesting. I've noticed the same thing with um especially some front end work where like I said for me I'm not the person to design UX in terms of flow and things like that I feel like I'm okay at but the design no. Um, but I've definitely asked LLMs like whether it's curs or co-pilot, whatever it happens to be. I like and I could leave it very high level. We need a search bar, we need this capability, whatever. And it does it. And I'm sitting there going, "Holy crap." Like >> I don't know how you did that. Like you said, it's because it's one shoted from somewhere, >> but it was able to make it look good, like feel good with like basically like not my input, which is super cool. Um, but you had mentioned too that like you're finding this balance and being able to like like really lean into how do you actually explain these things, right? How do I explain these concepts so that the LLM can go build what you're hoping that it will build? >> So, >> I mean, it's it's interesting. I mean, and I think it it also depends on what model um you're talking to. It's kind of like are you dealing with data models? Are you dealing with UI? Are you dealing with systems? Like I mean in the last 6 weeks I've really been pushing hard to see like I mean I mean I lit literally living in cloud code like probably 10 12 hours a day just just consistently doing things and it's like I'm writing very little code and I didn't really get that when people were like oh AI is writing 80% of your code like now I understand what they mean is because literally I'm just helping it along but it's taking the lead and I'm guiding it. And what I'm finding is I mean that like the the 0ero to one it's good at the oneshotting of things >> it sucks at refactoring like horrible horrible at refactoring and that's the part that is so painful where you let it write it so you're not super familiar with the code but when you have when you want to go and refactor and rework it it's hard because it feels foreign and that it's not your it's not really your code and you don't have it in your head. And so this whole thing seems like a sort of a caching problem where it's like, okay, how do you keep the current state of the code in your brain or in the LLM's context window so that you can answer things quickly? And um it's been a hell of a learning experience. Sam, >> that's a it's such an interesting point though. I was talking to someone recently and the way that they kind of frame that is that it seems like the models are very good at naturally like just adding more code. You need something cool. Let's go add let's go add. And like I think as you know people that have been writing code for a while if you if you're more experienced you know the feeling of being able to delete code and being like hell yeah like that's the best feeling like >> you know a thousand lines of red boom that feels awesome. But like the only time I see that happening is when the LLM screws up and it's like just deleting a file and I'm like uh >> but would you say that that's kind of a common pattern that with refactoring it's it's like it's not removing it's just like here's more. Well, and it I think it's it's inability to see patterns across multiple files is its biggest limitation right now where like right I was working on this LM streaming functionality for our chat and I wanted to reuse the basic and like we have a it's gotten to a point where it's a really nice like sentencebased streaming and like it comes in really nicely with a little cur blinking cursor at the end and then it finishes itself out. It's it's gotten kind of elegant and but I wanted to reuse that in another place in the app. And so I wanted to kind of refactor it and if it was me doing it, I'd be like, "Okay, I'm you kind of cash it all in your brain and think about these three different places that you're trying to kind of refactor and put together." it it struggled so hard and in and one of the things like it like to your your point of like it started adding things that I wasn't even talking about or it's just like it gets blinders on and it's like oh I'm just going to rebuild this other thing over here. Um and the other the other part that it's really bad it doesn't see negative patterns either. like it's like, "Oh, this one thing is actually different and it probably shouldn't have refactored it, but because I mentioned it in the in the conversation, it thought it had to do it and it just went and ripped out the code that was there before and essentially refactored it over top to make it look like this other thing." And then I'm like, "Wait a minute, like why did you touch this route and why did you touch this control?" And it's like, well, I mean, we were talking about refactoring and I basically just had to revert that that code and it just went off and got crazy. Um, but there's a whole other thread that I think is interesting of the model isn't the same model every day and and times of day even. And there's a great thread on Twitter today of people being like, I mean, did they nerf the like the claw model? And but I mean I've been finding that what time of day you're using it like obviously between Opus and Sonnet you get a different feel but since I've been so hardcore using it for the last 6 weeks I've noticed this pattern where like it it gets dumber like no joke there is a period of time during the day and even as of late it's dumber than it was two weeks ago or three four weeks ago. And I swear that when I started working on this project, it just felt magical. Like it was just so fast. It got me. I was having good conversations almost like a real pair programmer with it. And for some reason that has gone down over the last several weeks. >> And you genuinely feel like and I'm not challenging you on this. >> Yeah. No, no, please, please do. Yeah. >> But uh but you genuinely feel that because you said you spend a lot of time every day doing it. So that perspective that you're sharing is not like you didn't get acclimated to it or something. You genuinely feel like it it's like it's literally getting worse. >> I mean, a lot of time it's like I can tell just from my frustration level with it where I'm telling it to do something and it'll look left when it should look right. Like it's just like completely wrong. But it's overconfidence is the part that is frustrating where it'll be like, "Oh, yep. Fixed it. Okay, great job." Like whatever. and then you look and you're like you didn't understand anything I just told you and I just know that that wasn't happening as much at the start of like last month. Um and I think I mean I I definitely know I can tell when it drops from Opus to Sonnet like that's a pretty big jump. Um when you're like on cloud code I'm on the max plans so you like basically can use Opus for so much of the day and then it drops. Um, but there's just I mean there's there's something like you can even look back at my tweets from a month ago where I'm like this is crazy. Like I cannot believe how fast you can develop some of this stuff and the feeling of like communicating like it it was almost like I was talking about the architecture and it was like we were riffing in different ways like it was a really weird surreal like Westworld kind of feeling a month ago and I haven't had that lately at all and it's just and and it maybe is an acclamation thing but I also just know how many times it screws up >> and like that's a that's a tangible thing at this point. >> Yeah. And that's why I was curious cuz like that's why I said I didn't want to >> challenge you and make it sound like I don't believe you, but it was a very much a curiosity like cuz if you're using it that much, you probably would really pick up on like something's going a little south here. But um >> well then I mean this thread that I found today I mean there were hundreds of people saying similar kind of stuff and I was like holy crap like that's there's something and they said basically the last 10 or 11 days it's gotten different and so either they're turning down the compute like they're trying to spread the load more um something's going on like I don't think they're changing the model some people think that but I don't I think it's like there's literally just allocating less GPUs to it or something simple like that but there's something going on because I mean it's like I mean I've I've just been using it so much that I it's it's a weird spidey sense that you can tell. >> Yeah. And that's like I appreciate that part, right? Because if you had said like I only use it like an hour a week, I'd be like I don't know then. Um but when it's that much it's like yeah like you really pick up on patterns. Um super interesting. Um I don't know if you have like from your perspective uh sort of given how much you're using like claude code or other tools to be able to to to do programming this way. Do you have like general guidance now that you're kind of like setting up for yourself or you're like okay I got I want to go build a feature let's say um you have an idea in your mind and like is there like a set of steps you're like okay like this is my approach now like my set of tools and my approach that maybe you can share >> yeah I think it's interesting because I mean I don't consider myself a front-end dev like I don't like I wouldn't say I know react like I know the concepts like I know componentbased kind of layout and I know I I could like but like I couldn't pass like a front-end dev test right now and but I've built full UI by myself in the past like with ASP.NET and with other stuff and gotten deep into CSS and all that crap and like I mean but it's like right now I'm like okay I'm like an idiot soant where I basically know what I'm trying to ask it for but without doing a little more training and just getting my head around it I couldn't do it myself. And so I think it's interesting where I can typically give it like I always start with like a PRD style where I'm I riff with it for a while, talk about what I want, and then summarize it and ask it to to write it as a as a PRD. And that works pretty well. And so I just like throw a bunch of stuff at it. Maybe give it some competitor websites, give it um give it even uh screenshots, um some stuff like that, and then let it digest it. and summarize it into a PRD. And then I'll typically read over the PRD, maybe edit it, maybe riff on it a bit, and then I have it build an implementation plan. And that generally works pretty well. Um, it gives you a sense of is it on track, is it off track. Um, and then I mean really at that point then we just start chopping it up of okay, here's here's the different pieces. Um definitely I would say my suggestion would be try not to change too many files at once. Once you get above 5 to eight files that it's in the change set at that one point I think it you get diminishing returns. Like I've definitely screwed myself with like 20 files that it's trying to do at one time. It just can't keep track. Like >> it's like the context seems to just like completely fall apart. Okay. And I mean it's only 200k tokens, but you really with cloud code especially, you can't really see what's in the context window. Um, clin I like they do a nice little bar at the top of like kind of what's in your context window. Um, and cloud code gets horrible as it gets closer to the end of the context window. Like it's it's literally you could see its IQ drop like in proportion to um and not that's the crazy thing. It's like it is so bad sometimes like you'll you'll be like what is it doing? what is it saying? And then I'll look and it's like 1% of context window left and it's just the weirdest thing. And but I think that managing the context window, managing its edits, prepping it, keep more in cache. I mean it's such a simple thing of you just you want it to be talking in RAM with the RAM like with RAM almost in that sense but it um and but even though I mean what I found is tell it to tell like ask it to tell you about what it did. That's I think I'm sure this isn't new to a lot of people, but it's like that's my trick of like okay I mean walk me through it like say like promp me walk me through step by step why you made that change it or whatever and it it in one hand it causes it to think but in two you're essentially code reviewing what it's doing and you get a chance to like it's easier than just reading it. Um, and then I have one little command that I do called bet that is I basically ask it to bet $1,000, like up to $1,000 on it working 100% um, correct out of the box. And there's like only three or three sentences maybe in the in the prompt, but it causes it to go back and do analysis on what it missed, how the quality is, and it's a really simple prompt, but it works really, really well. That is fascinating. >> Yeah. And I just kind of I just came up with that, pulled it out of my ear, but it's like it it'll be like, "Oh, I'm only at 200300 bucks." And you'll be like, "Whoa, whoa, whoa." Like, "Where? Why? Why?" Like, "Do you think it's like you screwed up this hard?" Other times it'll be like, "Oh, no. I'm at like 750. Here's what I think I need to fix to get to 9 950 or whatever." Because I basically said like, "Try and get above 900. 950 is optimal. Don't iterate too far to get to a,000. like don't try and be perfect. And for some weird reason that that works pretty well like that whole that um process >> that is okay. Again, lots to unpack. The the first part that I thought was really interesting, I've heard other people saying this around maybe not the exact same way, but basically the you described it as having a conversation with it in the beginning, right? So you in your mind have a feature you want to go build. >> You spend this time literally having a conversation with it just to go back and forth. I'm assuming at this point it's not outputting code. Maybe if if anything it has maybe, you know, some pseudo code or something it's dropping in the chat, but you're really just talking back and forth about what you want to go build. And then before you have it go off to write code, it's like >> summarize it and we're going to review this basically. Yeah, >> I think about that and it kind of carried over to me awful lot like if you were to go work with another person and try to build something, >> if you were to just go up to another dev on the team, even if they're a really solid dev and you walked up and you said, "Build this feature," and you gave them like two sentences, they might in their mind go, "Hell yeah, I can go build this." They go build it. then you look at it after and you're like, "Nope, that's not what I >> like." We would fix that by having a conversation and like talking back and forth and, you know, writing down some ideas. Um, and again, however this looks and however formal you want to make it, it doesn't mean you have to go write a 10-page design doc, but >> the point is that you have this conversation, this review process up front where you go, >> do we feel good about this? And then I love your your your take on this where it's like to to bet money on like like how confident you are. Um and I guess regardless of how you approach that, but basically forcing it to do its own analysis. >> Um >> which I think >> some of the things I've been trying out like maybe not to that degree are like I might use um maybe I'm I'm just going to use an example. I might be talking to chat GBT doing this riffing kind of idea >> and then have something where I'm like okay I want to go I didn't realize I was talking to chat GPBT and it's not an agent so I have to go like now I want to put this in cursor in co-pilot so I will say like basically summarize or write this in a way that an LLM can go execute it like step by step um so not the exact same thing but sort of using LLM to summarize feed into other LLM is kind of interesting >> and and I think I It's I think to your one point before is like I don't know why I did this but I just I treated it like a chat session like it was like it was like I'm DMing with somebody in Slack or like I'm talking in Slack or I'm on like I'm in a bulletin board or something the old day like it just I just treated it like it was a person and I mean not overblown but just like I just talked to it that way and that for some reason just works. So, I didn't I wasn't like using it like a um hey, go write this for me and like very like robotic kind of interaction. It was really more like, hey, I'm thinking about this and and I just I just started just like talking to it. And I think it at this point the models are good enough that they can they can emulate that flow. And I think I didn't really see until the recent generation at least with like 3.7 of um and four of of sonnet and stuff like they could really work well. And I I think I tend to like the anthropic models anyway just for >> there's something about their their I don't know you want to call it personality but just like the flow and like they're they're a little more chatty sometimes like they just they're more and I so I think it's a I I tend to lean more towards anthropic in that way but um so I just kind of rifted with it and just saw it was like okay well what can we do here and that was where the first week I was just like it was crazy because I would even bounce ideas for architecture off it and be like what if we did And it was like, "Oh yeah, like that's a I mean, oh yeah, you could do this. Oh, that'll make a great feature for like it it got the company of what we were doing and the product and I was actually working with it writing a pitch memo." And so it like it really got what we did. And what the funny thing was is I then saved those in markdown and whenever I wanted to refresh it on what the product we were building, I would just be like, "Hey, read the pitch memo." And I think that was that was just a grounding kind of thing for it. And it's it's great to have access to that kind of content. That's a like a a super interesting and sort of unique like opportunity where you can kind of snapshot what like you you were kind of talking about this a little bit earlier about like you know how do we get a snapshot or like a memory of what it's doing and like being able to say that well this is a good conversation or you're using it for something and put it in markdown now we can leverage it later. There's um I'm certainly not like you know cutting edge on agent uh development. I'm trying to experiment with and I have no idea if this is actually helping or not but for my own projects I don't write documentation because I'm living in breathing in the code. I love to work in it. It's fun. Um but writing documentation I don't like but I've noticed that with agents well the first thing I noticed is that it just loved to start adding markdown files all over the place. >> And I was like get rid of this get rid of this. And then I started thinking like wait a second if it's writing documentation as long as I can make it such that it's not like uh documenting here is what this change does like it's a transient kind of thing I want it to be like let me document the feature this is the architecture this is why this feature exists not a >> I don't know the right way to say that but more permanent not transient >> I started wondering if I have this stuff more in my code base is that going to help it so that when it's trying to do more work in these areas in the future or if I'm like hey there's this architectural pattern go X that it's like hey there's a readme there or there's some other markdown file I can pick up on this >> I don't do you have like have you seen that work because I'm trying it I still don't know if it is >> I mean I think the one thing I I'd heard about everybody talking about like cursor rules and clawed MD file and all that kind of stuff and the one problem is it's not obvious when it reads it like at least it's I'm always like, wait, I start a new session. Did it actually read the cloud MD file? Like, do I have to tell it to? Um, and so I think there's some of that is like I basically um I definitely have found that having it read like like now I was working on this refactoring. I created a refactoring plan or had it write a refactoring plan after we talked about it. And so now it's it's partly just because the context window you got to kind of reboot it every once in a while. and I just go and have it re reread that plan. And I think for that it's like it's essentially you're paging its memory to disk in a summarized way and you're asking it to reload it before you start again. >> And that tends to work really nicely. Um, but I think I mean it's not going to go the problem is it's like at least with cloud code it doesn't think about going to search for stuff agentically. Like it really is kind of dumb. Like and I know with some of the other agents like some of the other agents are a lot smarter about being like, "Oh, I'm looking at this file and I'm going to figure out what's related to this file. I'm going to go look for stuff." like it just starts gpping and like searching and trying to find stuff from ground zero every time. And I think that's where like there's such green field for I mean I don't know if it's going to be a graph rag. I don't know if it's going to be whatever but it's something and I know that a lot of people are like oh no you don't want to index this. It's not semantic search, but like you got to do something like you can't look at a file >> better than that >> or or or just I've been thinking about like can you just leave breadcrumbs around like of okay you're in this file it's like here's other three other things that you probably are going to need as well and maybe that's where having a read me per folder or something like that and it like if you've seen what Devon did with um what's it called deep wiki like how they they went and researched all these GitHub repos and they have these summary stories of like all these repos like something like that where you essentially create like a markdown index like of in this folder these are my Zustan stores here's what each one does like and but then I think it it needs to think through how to go find other stuff better um cuz I that's the part that just gets tiring is like oh I closed my context window now I have to go tell it where all the code is again and like I can't that makes us feel like we're like in the in the 70s still. Like it's just seems so old school. >> Well, and this is the the one more thought that I had before I wanted to kind of come back to to talking about your your platform again, but uh as you're talking about this, it feels like to me it feels kind of ridiculous that it's like this is obviously very powerful. Like without a doubt, this is ridiculously powerful technology that we have to work with. But some of the problems that we're facing, it's like, you know, like why are we gpping through things? like why do we have to like you said we don't even know if it's actually going and picking up these MD files again. >> Um it just seems like these >> like they're very ridiculous problems. Uh and don't get me wrong I understand that this is it is complicated but they seem so silly given how powerful it is. >> Yeah. Well, I mean, and well, the thing is like I came from a Visual Studio world, not Visual Studio Code, but a Visual Studio where like it has great IntelliSense and it has great like background compilation and like the the turnaround time working in C because that's kind of my main my main language >> for 20 20 years now is like that's my go-to and moving to do Typescript. I'm just like this is I mean I could go off and I don't want to bash TypeScript but it's like it's just like the speed of development is so much slower and it makes me just be like wait is this how people work normally like it takes this long to lint a file or like it how the the type cache doesn't catch up with showing you what like is just changed in a background comp like and that's the kind of stuff was more of a net versus TypeScript world that I had never really built anything with with TypeScript other than little sample apps and stuff like that before. So I think this is where and but the crazy thing is the LMS aren't that great at it. Like they make mistakes all the time that lint catches or things with React dependencies and stuff like that. And that's the crazy part. I'm like is it just like is it just trained on bad code or is it are people just writing bad code? like what it's it seems like a training set problem with a lot of the the and I wanted to try the V0ero model that those guys built that I think was more specifically trained for like Nex.js and React. Um, but I mean I would almost love to be like, okay, Opus 4 like with a I mean my my perfect world would be Opus 4 with a million token context window fine-tuned on TypeScript and React >> like would would be incredible and then I could really get some work done. >> Yeah. No, it's it's super cool. like this this stuff is advancing so fast and I I can't wait to like in 10 years like look back on conversations like this and be like we were really dealing with like we were like we had context window issues and like it didn't know to like go look for this like that was so ridiculous >> it reminds me of literally like the the old 640kb Windows days and having to like page out to like get virtual memory like it's I It's exactly the same thing. I mean, it's literally the same patterns. And so, I mean, it'll happen. I'm not That's the thing. I've been around this long enough that like it'll happen, but it's funny looking at it now and being like, okay, like some of it's magic, some of it's just a pain in the butt. And so, >> Right. >> Well, I I wanted to say thanks again for sitting down to talk with me, Kirk. And, uh, again, for for you have a a launch that's coming up. You're hoping by the end of this week. you want to kind of walk us through again like what you're going to be launching and and where people can find out about your product and service? >> Yeah, so the the platform API is called graphlet. So I mean graph lit.com um so you can get information and sign up and the API docs are there but the product we're launching this week is called Zen like magazine so ze um.ai AI. And so it's kind of based off of like back in college days, everybody would kind of build their own magazines called Z. You could like build on the photocopier and tear like sort of cobble together like an unofficial kind of publishing model. And so that was kind of the riff of what what Z. And it's I mean it's going to be free to sign up. Um so there's a free tier. You can upload data like it has integrated web crawling. You can um pull in YouTube audio. um all those kind of things as well as um on the page tiers you can get your slack, your notion, your Google drive files and create context that is um organized. So it's like if you're thinking in chat GPT now you can connect your data in there but where's the sort of photo like interface to see what you uploaded? That's what we're solving. And so we're a full context management platform but then it's all about what can you do with it? um publish that data out, summarize it, like search for it on a map um and those kind of things. And so um hope hopefully people find it really useful. I think it it covers a lot of domains. Um even personal use like go search through your email, look for um look for bills, try and put it on a timeline of what bills are coming up, like you could build workflows and things like that. And so um but yeah, so Zen.ai um there's a wait list right now, but it'll be out um this week hopefully. And then um anybody can try it. >> That's super cool. And I mean to your point and you were saying like personal use, right? As you're talking through this, I'm like as someone who creates content on online, I'm like I have tons of YouTube videos. I have tons of vlogs where it's just me talking and like ingest all that information. I need to make social media posts. I'd love to be able to say like get this content narrowed down to some topic or some context window and go, you know, basically use that data to go create more content. like I I just feel like >> I'm giving you one very specific use case and I'm sure there's tons and tons of other ones. That's super cool. So, um >> yeah. No, the public I think the publishing is the thing that is really fun where you can sort of curate a set of data and then say, "Okay, go create me like a blog post about it or go write me an email or even create me an audio summary." And we just have like little one-click kind of things. You select your stuff, you pick your template, you can even customize your pumps, hit go and we basically do all that work and come back and and the nice thing about it is we reingest the file at the end. So you can then it becomes searchable and so now it becomes a derivative work that you can keep working with. Um and even we we added in the like the image gen models from OpenAI. So you can even create little infographics, little ads knock and and that's actually super fun. You can literally just upload like a one pager of your product or something, your web page, and be like, "Hey, go create me an ad." And it'll just do it. And that kind of stuff I is just super super fun. >> That's awesome. I'm I'm personally going to check this out because I have not yet. So, um I am excited to check that. So, um I'll double check with you after and I'll make sure I get links and stuff in the description. And uh Kirk, thanks again. This is super cool. And I have some things that I'm taking away for my my agentic coding as well. So thank you. >> No, and I think I mean the long term it we want to then become integrated with like other agentic frameworks. So it's I mean even Zen might be the way that you curate data to go use in your MRA or Agno agents and things like that. So we want to have it kind of be full circle. So it's it's exciting. I'm really the the rest of this year is going to be interesting to see how all these pieces fit together. Well, and that's why when you brought up MCP, right, like being able to be on both ends of that is like really powerful because now you have systems where you can be, like you said, going full circle with it. So, it's super cool. >> Yeah. No, I'm I'm I mean, honestly, it's like the most excited I've been about a product launch in the in the company, and I think it's it uh I'm really hoping it it uh gets some legs. >> Well, it sounds like it will. Uh thanks again, and I'm definitely going to check it out. So, thank you. >> Okay. Oh, thanks for thanks for the time. Of >> course.