Scott and Mark Learn to Vibe Check

Heat. Heat. Heat up here. Heat. Heat. Heat. Heat. Heat. Heat. N. Heat. Hey, heat. Hey, heat. One night. Hey. Hey. Hey. Heat. Heat. N. What about in my Hey, I'm What are you? What are you in? One night, one night. night. Are you Hey friends, I'm Scott Hansselman and I'm here with Mark Vinovich who has absolutely no idea what is going on today. Have you been prepared or prepped in any way? Zero. Okay. Is your mic even on? There you is. Is it? It is on. So Mark does not know what's going on. And this is exactly how I like it because he makes me uncomfortable on our podcast. Scott and Mark learn to uh every week. So now it's our turn to make him uncomfortable because this is Scott and Mark learn to vibe check. So here's the deal. Nobody likes AI slop. Nobody wants an AI to create applications that are not awesome. So we think that there are three kinds of applications. There's AI slop, there's vibes, and then there's AI engineering, proper AI augmented software engineering, which you're a huge fan of. Definitely. So, what we've got here, my friends, is four wonderful creators from out in the community who have created applications and in the style of pen and teller fool us. They're going to come up here, present their applications, show us how they made them. They're going to show us their code. We're going to judge the applications and we're going to guess is it slop, is it vibes, or is it AI software engineering? And it's going to be awesome. And the points don't matter. And we can ask them questions about how the code works. Exactly. We're going to ask them questions. We're basically doing live interviews on how they made their stuff because we want to make sure that everyone is doing great work with things like Copilot, with things like Claude, with things like Codeex. Can we make quality software or are we going to be gaslit by the community? Are these people going to fool us? And we're going to find out. It's going to be pretty cool. We also have to thank our sponsors. We've got a lot of folks that have put a lot of work into making things like this happen. When we have crazy ideas like this, it's our sponsors that make us uh possible. So, we're going to shout out to our sponsors and we're going to come right back with our first guest. Here we are at Microsoft Build in San Francisco and I'm coming to you from the wonderful AMD booth right here on the show floor. I'm really excited to talk to Adrian because there's so many fantastic things that are happening to really empower our entire developer community right now. Adrien, welcome. Yeah, thank you. I'm excited to talk about it. There's so many things happening in this space. I'm sure we could talk for a long time, but let's cover a few highlights. Let's do that. Let's What's really important about the partnership? Why does the partnership between Microsoft and AMD matter to this community? Well, we're going to talk a lot about the technical uh innovations that are happening in the space, but tools and technology isn't enough without a strong ecosystem. And that's what we're working on together with Microsoft is building a community that's going to nurture and foster that innovation. It seems like there's this crazy pace of change. It has completely altered the way that we think about work and work forms and these platforms as a whole. Where I know experimentation is something you're passionate about. Tell us a little bit about that and how we should be thinking about that. I think that's the most exciting point here at this conference today is that the rate of experimentation has really blown up. People have so many more choices. Not just the technologies they're choosing, but the capabilities and models and platforms. Uh the choice of where things run, how well they run, and what runs. All of these are new experimental choices that design designers have to explore with. That's amazing. And there's a cost impact there as well, isn't there? Well, the cost function is something that they now can control more readily. Before they had to make hard commitments on the choices they were making. They had to plan for them. They had to commit to those decisions. Now the cost to explore is zero. Really that is so empowering to people. It can be a little overwhelming too. But I'm excited by the opportunity that is happening there. And I know that there's going to be more that you're going to be talking about here. There's another conversation and there's assets. So I just know you have so much to offer. Thank you for the partnership that you have. Uh this is fantastic. It's such exciting stuff. Now, if you really want to dig deep into everything that is being talked about here, you're going to want to visit AMD's sponsor page on build.microsoft.com. That's where you'll find all the resources, links to the sessions, and everything so that you can get busy right away. All right, make sure you do that and I'll see you again soon. Hey friends, we are back. Thank you so much again to our sponsors for making this happen. This is Scott and Mark. Learn to vibe check. Our first guest today is the wonderful Steve Sanderson. Steve Sanderson and is extremely well-known programmer. You've probably heard of things like KnockoutJS. He's worked on Blazer. He's done amazing things on the internet. And he has brought us something today that we have not seen. No one here has seen this before. We don't know if it's completely bespoke. We don't know if he wrote it in a caffeinated co-pilot session. It could be slop. It could be vibes. It could be AI augmented software engineering. What do you have for us, sir? All right. Yes. Um, so what I'm going to show you today could be pretty momentous. So I need to start by putting it in slightly historical context. All right. So the human journey, right? We've come a long way. Stone tools, language, and culture, mathematics, computing, and these things are all quite good, but I don't think we've really reached the pinnacle of what we could be as a species yet. or at least until what I'm about to show you just now. Okay. So, I know you're probably all thinking like, "Oh, he's going to reinvent the whole software industry and we're not ready for all this." I I know you've all had a lot of change to deal with recently. You know, we've moved from this world where humans write code into a world where AI writes code. And I know that that probably feels like quite a big deal to Mark and Scott, but for a true visionary like myself, I'm already thinking about where we're going next. And in fact, I'm already there. So, I'm going to show you and I need to introduce you to a completely new operating system that's going to change everything about how we work. So, introducing Vibe OS. Okay. It's the world's first completely hallucinated operating system. All right. And I'm going to show it to you right now. So, firstly, it is a real operating system. I can boot it. It's a VHDX file here about 1.3 GB. Could probably make it smaller. I didn't bother. Um I could boot it on bare metal, but I'm not going to. Going to boot it in HyperV. Uh and so you will see that boot up nice and fast. It's very efficient. And when that comes up, you'll see it's got a beautiful user interface that allows our user to be highly productive. Okay, so here we go. We're ready to get started with that. It's a bit hard to see there. So let's just switch to a full screen view. And um we're going to start by running a few of the applications that are built in here. So we'll get some of the classics going. So we'll have Notepad, we'll have calculator, we'll have Internet Explorer. And the thing that's unique here, the thing that you will never have seen before is that in these applications, there is no code at all. Everything in all these applications is being hallucinated in real time by the AI. So there's no buttons, there's no event handlers, there's no logic to say what to do, but the application works. So if I do five divided by three and then uh equals, if I can find that, it's up there this time. Um, it's 1.6. Okay, so it works, right? But I didn't write any code and neither did AI. There's no AI written code either. It's just producing UI exclusively with nothing behind it. Uh, and we've got Internet Explorer here. Um, unlike the boring operating systems that you use today, this solves one of the big problems that people have, right? People are always wondering when they're on they're on their internet, is it AI or not? But in ViOS, you don't have that problem. There's no question to be answered because the answer is always yes. Everything is AI. So, if I want to, I can go to, I don't know. Let's see if we can find uh whether Scott Hansselman has a Wikipedia page, shall we? So, I'm going to go to uh google.com here and let's do a search for Hanselman Wikipedia and we'll do a Google search there. And I don't need to remind you this, but all of the content you're seeing is completely hallucinated, right? So, it's just coming directly out of the model and the UI is being updated in real time as we go. And here's a beautiful picture of Scott Hanselman. I'm not sure about that one. All right. But it goes further, right? We're not limited to these built-in applications. We can search for any app that we want. And everyone always does like to do app demos. So, anything that I search for, I'm going to find, and it will just be created for me on the fly. All right. So, let's see what else we could find in here. Um, let's go for one of the classics. Uh, En Carter 98. Everyone loves that one. But it's all about Mark Russino. All right. And so we can get a fully customized version of En Carter 98 on whatever subject we're interested in. All right. So lots of facts and figures about um Mark here. Let's see. Selected facts. Civic has a talent for turning internals into folklore. Great. I assume that's true. I don't know. All right. and other applications that you would want but normally it's too difficult to make. So like commander XCE but it's always rude to you. All right so that's the sort of application that I think is a genius idea but it's difficult to get someone to implement something like that. Uh so let's see what we've got on this machine here. And obviously this is always different every time. Okay there's something suspicious. Let's see if we can run bash in here. What's it got to say to us? Uh bash uh you have somehow offended. Okay, fine. All right. So, you get the idea, right? This is a pretty revolutionary way to do computing because we don't need to write any code for any any of our applications anymore. Is are there any software you would like to to try out within Vibos? Anything you can think of? Uh Microsoft Money95. Money95. Civilization peaked at Microsoft Money 95. Uh but but Scott Hansman has a lot of money. Oh, but for Scott and Soulman. Yeah, you can say Scoot as well. That's totally fine. All right. So, let's see what uh money Microsoft what Scott had in 195. Uh, okay. Let's see. Okay. And this is this is Scott checking account. What model is in the back end of this? Scott, um, it's been a tough month for you, I guess. Where where are the tacos? There we go. That is amazing. All right. So, all right. I want to see paint with a drawing of Scott on it. Paint with a drawing of Scott, though. Uh, let's do for your haircut, but with pre-installed. Okay. And so, and you've got this is Copilot uh SDK that's doing this. This is using Copilot SDK, believe it or not, behind the scenes. I can show you a little bit about how it works if you want. Um, so it says it's got a very normal picture of Scott Hanselman for some reason. I don't know what it means by very normal. Um, okay. There's a normal picture of Scott Hanselman. That's your cage. That my cage. That's the cage you keep me in. Okay. This is insane. Um, wow. Okay. Could it This is This is Vibe OS. Can it have nested OSs? How many How many OS deep can we go? Can you simulate like a an altar uh 8080 that's like a terminal emulator? So I need a terminal emulator going shushing into a altter. Totally. So we can get iOS simulator in there. We can What else could we have? We could have um this may be the greatest thing I've ever seen in my life. Me too. Oh yeah. Well, fair enough. All right. So uh what have we got? Oh, it's difficult to see on this screen size, but yes, we got a fully functional version. Right. And I love that it uses the the appropriate Windows CE buttons that you are likely to see in an iPhone. Yeah, it's very very legitimate. So, okay. So, did he write this from scratch or did he vibing vibes? I My mind is blown by this. I know. I told you it'd be fun. You have questions. Please, you have three minutes. Define hallucination. I'm sorry. I define hallucination. You call this a hallucinated operating system, right? It's hallucinating everything. Are there sound people able to make the sound come this way as well? I really can't. Louder. Can you hear this? Not really. Um he says you're define hallucination for you. Define hallucination. Okay. So, um basically, all right, let me show you the code. Right. So, behind the scenes, um what we've got is inside our editor. Right. So, what's happening here? I went through a different a few different ways to try and do this. First way I tried to do it was by just generating raw bit map images using like stable diffusion that kind of approach and it was sort of able to produce images of the desktop operating system but it was rubbish because it took ages and when you click on it all you can say is like the user has clicked at these coordinates and it doesn't really know how to update things. So then I switched over to generating a structured representation of UI. I tried a few different versions of it. I made a custom DSL. I tried this thing called cute which is a native um UI thing and I could generate a DSL for that but it didn't do a very good job. The thing that really suddenly was able to do a great job was when I switched it to just b plain HTML because you know these models have seen so much HTML they can make that up very very easily. So what we're doing here is every one of these windows is a separate iframe and when it pops open each one of them creates a separate copilot session within copilot SDK and it gives it instructions like this. So it's saying you're simulating this application UI. I want you to produce some HTML that represents it and it advises the model to put IDs inside all of the elements. And then whenever the user clicks anything, all we do is we send a message back to the AI saying the user has clicked the element with ID R1 or whatever now produce a diff that I can apply to the HTML. And so then the model produces the diff, we apply it and the UI updates and you get this kind of like fake statefulness because it's all happening within one copilot session there. So yeah, see the irony here and this is where Steve is very clever is that he is using his AI augmented software engineering abilities and skills to create slop. And that's the genius. It's it's brilliant. It's absolutely brilliant. I love it. I am intentional slop. Intentional slop. But only one that an AI engineer could possibly do. Yeah. Do you think you could have done this? Do I think How long do you think this would take you to do? So, this vibe coded thing. How long would it have taken? A minute. It's a while. This is a This is a few hours. Couple hours. How long did this take you, sir? Uh, it took a few evenings. So, yeah. I had a I played with a few different approaches. Yeah. So, total total time, you think? And you were probably multitasking while you did it. Yeah. Total time maybe five, six hours or something like that. Five, six hours. I would say this is two notches past vibes heading towards AI augmented software engineering. I agree. Yeah. All right. Fantastic. We're going to give you an award. We brought you an award. I'm going to get up and we're bringing you this award. Thank you so much. Congratulations. That was some amazing vibes. We are going to go to our sponsors right now, but shout out to Steve Sanderson for an amazing opener. Wasn't that cool? That was big hand for Steve, everybody. All right, we'll head out to our sponsors and we'll be right back. Thank you, sir. Welcome back to the show floor at Microsoft Build. I'm so excited to be here at the beautiful Intel booth and I'm here to have a conversation with Brian, my friend Brian here who is going to tell us all about what Microsoft and Intel are doing together. Brian, welcome to Microsoft Build. Thank you very much. I hear it's your first one. First time. Well, we're happy to have you here for sure. What do people need to know about the partnership that we have between our two companies? Yeah, I mean Intel and Microsoft have a decadesl long partnership and we've been supporting the developer community for Intel has for almost 30 years now. So it's been a very long time and we're very proud of that and you know one thing that we try to do is we try to you know help developers to optimize their applications across the compute continuum and what I mean by that is you know not every workload is is optimized for a specific compute type. So there could be CPU, there could be GPU, it could be NPU. So we are helping developers to optimize their code to run best on the right infrastructure for their workload. Oh, that's good. How do they get started with that? What do they need to know to move forward? Yeah. So, you know, we have 100 million AIPCs, actually more than 100 million AIPCs out in the market that are that are made to run on uh or capable of running across those those compute infrastructures. So, we also have uh software toolkits to help developers to optimize their code to run best on the the infrastructure that's best for that workload. Oh, that is fantastic. Now, everyone's always really concerned about security. What about that? Absolutely. I think security, especially in this kind of age of AI, is growing in importance and you know, Intel has developed a technology called TDX which enables confidential computing in Azure. So, essentially what that does is it allows developers to run sensitive data and sensitive workloads in a confidential computing environment and have the peace of mind that their data is secure. That's fantastic. You can learn about that and more on the partner showcase page. Make sure to check it out. Whether you're with us in person or watching online, sponsors elevate your experience and help accelerate your goals. Explore how Enthropic extends coding agents across your team with clog code and co-work in Microsoft Foundry. Combine powerful infrastructure, data, and applications to help your organization innovate faster with Oracle. Find out more about how our sponsors are using AI to solve the issues of today and accelerate the innovations of tomorrow. All you have to do is visit our sponsor directory for the latest advancements, inspiring customer solutions, and special product offers. Sponsors make more possible. And we are back with Mark and Scott learn to vibe check. We have just seen Steve Sanderson make a vibecoded entire operating system in a couple of evenings with the power of AI. Amazing. It was amazing. You were impressed. I was impressed. It's when I when I see a project and go I wish I'd thought of that. That I love that. Yeah. Everything Steve does I'm like I wish I did that. But now today we have Cassidy Williams who is going to share with us what she has made. We have no idea if it is AI slop, if it is vibes, or if it is AI augmented software engineering. Cassidy, what do you have for us? Hello everyone. I have been thinking a whole lot about databases lately. There's so many new ones coming out, so many things. And so I thought it'd be interesting to build one in the browser where it just runs in the tab. And so if you can see my screen here, this is a little query shell. And so right now it's doing select star from ideas. This is a query that you've probably seen before with just this little table here. I could change this to say for example select title and then uh I have a ranking in here called cursedness cursedness from ideas and then it will filter based on that. I could also add a wear clause where I do where curs cursedness is equal to catastrophic and then I can run that query and it'll just work. And so I can continue adding things. I also had a query builder here so I wouldn't have to keep typing things. So cursedness could be medium. I want to filter by certain things. I want it to descend. I want to limit things. I can do all of that. And then once I minimize this, I can run the query and it will show. I could even add a drop table in there and then it'll say, "Oh no, this is loadbearing." So this is all very fun. I'm also able to attach a sample SQLite database in here and then if I just run the query again to select everything and not drop that table, it will load. And that's great. You've all seen this before, right? Where you can see a table. It's using web assembly to pull in the SQLite DB in here. But there's something that's a little bit more fun about this. There is no database and there is no query engine. It's done with pure CSS. Oh no. Exactly. Oh no. All of the querying being done is in pure CSS where yes, I use JavaScript to actually load this onto the page and it's using HTML to show the things. But all of the queries that you're seeing is done with just CSS selectors, CSS variables, if functions, all of that. Should you do this? Probably not, but I did. And that is what I did here. Why would you do this? I don't know. I I was like, what if I built something for this show specifically where I was like, yeah, this is neat and what if I made it terrible? And that's kind of where it came from. That's exactly the spirit that I was looking for for this session. Good. Uh, do you know CSS? Oh, I love CSS. CSS is one of my favorite programming languages to use and it is a programming language. When you clicked a chat SQLite and that dialogue popped up, does that file do anything or is it empty? Is that a real database? That is a this is a real data. Like the thing that you're seeing on the table is real SQite that is loaded on. So you're literally querying I'm literally quite an I say SQLite tomato. That's correct. Does anyone know SQLite? SQLite. Anyway, it's it does that all again JavaScript and web assembly loaded onto the page and then CSS does everything else. All of these check boxes are using CSS to query and filter things. I want to see the code. You want to see the code? I got you check it out. You need to make the code a lot bigger. I have read to make the code a lot bigger. Okay, one second. Let me pull this up so you can see all the not disrespectfully large. just click. Okay, cool. And I made it light mode so that way it can be seen, but I'm sorry if you're going blind. Okay, so this is truly just HTML. There's nothing here. You could see all of the CSS or all all of the JavaScript loaded it on the page and everything else is in CSS. The Chrome CSS file is truly just for styling that HTML so that way it's pretty. I picked a random theme in there, but everything else is done in CSS. So this table right here, let me hide the sidebar. It pulls in everything as a grid, but then it uses the CSS variable sibling index and that is used specifically for limiting. For example, every single query has a CSS variable in there. And let's see. I have Did you And you know CSS? You're one of the people that understands it. Oh, I know CSS. Okay. And when you were do you use you use GitHub Copilot? Yes. And uh did you use the Copilot app or do you use the CLI? I you I so I used the GitHub Copilot app uh alongside writing things in VS Code and every single time I was like this is gross. That's where I uh used some some assistance. Okay. How much writing did you do versus just yapping to the machine? I'm sorry. How much actual writing of the code did you do? How much is bespoke? How much of this is bespoke? Like like handwritten by me personally, right? Like you're like he doesn't understand CSS like Cassidy understands CSS. I'd say like 20%. You hand wrote 20%. Why? I like CSS. It's a nice language. You're one of the people that likes it. I there you might be the only one there. I see some people in the audience who I know like CSS. I know some faces. Yeah, it's a great thing. I love it. There are people literally nodding their heads next to people that are like nodding their heads. Come on. It's nice. Okay, so 20% of this you wrote yourself. Did you steer the AI or did you jump in and say, "No, no, no. I'm going to do this part." There were so many parts where I actually had to stop it where I was just like, "This button is terrible. This link should go here. some of these checkboxes aren't doing what I did. And so that that's where I would say, okay, I'm going to take care of this and then I would say fix. So faster for you to go in and manually fix than steer the AI. Yeah, it was a lot of human steering, human ideiation because I don't think any AI would say you should make a watching the thinking trace to watch it what it was doing or just look Were you watching the thinking trace to see what it was doing or did you just look at the and then you would stop it? I like having it go side by side. So I do have the GitHub copilot app where I would do some things, but then usually I like using the side chat so that way I can see what it's doing and have a bit more control. Do you think that we Mark and I who we know CSS like I know that if you say, you know, bang important, it means it's CSS, but you really mean it. I didn't know that. Did you know about important? Yeah. It's one of those things you're never supposed to do in CSS, but whenever anything doesn't work, you just say no. Seriously, what happens? You override everything. Does CSS is supposed to cascade correctly? Everything matters. You know, I'm more important than all of you other rules that important. Do you have any importance in here? I'm so sorry. Do you do any bang important? Do you break any rules? No. No. She I don't believe in those. She's a purist. Show me the part that you are the least proud of. The part that I'm proud of. Yeah. Of course. No. The least proud of. The part that you think is cursed. Oh, the part that I think the whole thing is. To be clear, no one should do this. I mean, this right here is pretty bad. This is terrible. So, this is this is a very like nitty-gritty behind the scenes. I limited it to 12. You you can't add more uh more than 12 queries into one thing because well, you could copy paste that line 63 and you could do 13. That would be truly cursed. Watch. Here we go. We're going to add what UI control you're doing. You're going to type it yourself. I do know how to type, believe it or not. Um, it does not want me to do that. It's starting to do display none. There we go. Look at that. Yeah, we've now extended this application. It It does. We have added a feature. Oh yeah, this is mostly mine now. I'm involved. Okay, I'll add you as a contributor. I'll add a thanks in the read me. Okay. And which part are you the most proud of? Which part? Show me something you wrote yourself. Ordering. Let me find it. So, one of the things that I was ideulating on a bunch was the ascending and descending. And the big fun debate and I need to find it cuz I admit I did some cleaning up was ordering things with flexbox. Do you guys know flexbox? I know you said you know important but flexbox is I know that the maximum number of nested tables you can have in net Netscape 4.0 is 32. Great. So I do not know what flexbox that is different. So, Flexbox allows you to rearrange things on a page and center things, put things at the front or back or anything for ordering by ascending and descending, which I don't know where my window just went. That's using flexbox and just using the browser engine in there to flip the order of things. Oh, wow. Okay. So, you not only used CSS, but you abused parts of HTML. Yes, that's what I'm saying. It's it's really not something you should do. And that's what's great. Did the AI push back and say, "Really, this is a bad idea?" No, it just said, "You're absolutely right." I was like, you know, I Okay. How long did this take you? Um, when did you message me? You You first messaged me about this. I texted you like a week ago. Yeah. You told me about this like like week and a half ago or something. So, yeah, I started work right then. But like clock time. Clock time. probably like two hours. That's two hours too long, isn't it? Probably. Okay. Is this Is this slop? Is this a The thing is with both of the folks, if we look at Steve Sanderson, you would need to be an expert to do what Steve did. You have to know and love CSS to do. Do you think you have to know CSS to create this app? You do. I mean, you have to know enough CSS to know that you shouldn't do this, but yeah, you need to know CSS to at least to at least if I were to not know anything, I wouldn't know how to massage it in the right ways to make certain things happen. What do you I think this is it's towards the engineered side of the getting towards AI augmented engineering and it's like you can't just vibe this. You couldn't oneshot this. How many turns were there? Yeah, you can say create a CSS based database engine in the browser. It just uses CSS. It's fully like in the tab. There's no back end. It's it's just running that and have it run off and come back and All right. Vibes, but it's two two or three notches to the right. Past vibes. Yeah, I agree. Fantastic. Cassie Williams, we have an award for you. Yay. Best vibes. We appreciate you. Thank you so much. That is cursed. Don't ever show anyone that again. It's open source on my GitHub. And it's open source on our GitHub. You need to add Scott as a contributor. Yeah. Yeah. Because I have a commit now. All right. That's fantastic. Thank you so much for your hard work for me. Big thanks to our sponsors. We're going to go to a quick break and we will be right back with more cursed wonderful things that we are going to vibe check. Hello, Microsoft Build and welcome. Boy, do I have a treat for you. I have Stephen with me from Nvidia and he's going to share some of the magic of the things that he's been building. And Stephen, welcome. Thank you, Carana. I appreciate it. We're so glad you're here. Tell our audience a little bit about what you have been building. I know we've had so much fantastic information going on about our partnership together between Microsoft and Nvidia, but you have something special. Absolutely. So, um, we have been building lots of things together between Nvidia and Microsoft, but you'll see a few special things here at Build. And one thing that I have been working on in particular is an integration with hosted agents, uh, Neotron and, uh, Hermes agent. So, this is going to be on display as part of a talk from Joey Conway. So, I highly encourage you to Well, that sounds fantastic. I always love to see the demos and the technology that people actually build. I think that's so important. Give us one uh important thing about this new implementation and why it matters. Absolutely. Well, it solves a lot of the problems that we see customers have in their Agentic deployments. Um number one is the control that they have over their agentic environments. Um hosted agents brings a lots of opportunity for customization and Neotron comes in with the high intelligence open-source um high capability model that brings that strong intelligence to the to the table and then Hermes agent wraps it all up and helps orchestrate your models in a way that brings the most value out of them. So, this really is, you know, a one, two, three punch um with these, you know, technologies al together. So, it's a lot of exciting uh new things and I'm excited to show it off. Fantastic. Well, everybody can see it at the NVIDIA showcase page and they should check that out right now. Thanks for joining me, Stephen. I'll see you again very soon. Thank you, Caruana. See you soon. Hey friends, we are back at Scott and Mark learn to vibe check at Microsoft Build on the big stage. Right now we have seen a vibe coded OS written by Steve Sanderson. We've seen a cursed CSS SQLite database written by Cassidy Williams. And now we have Swix. Sean, also known as Swix, who is very well known for his AI engineer world's fair. A great AI engineer. We're expecting big things from Swix here to see if he can pull ahead of Steve and Cassidy. What do you have for us, sir? Hey, Scott. Hey, Mark. Uh, hey, everyone. Um, I always wanted to do one of these like, you know, like Hello Sharks. Like, what if there was an app on the market, you know? Anyway, um, I don't have anything as fun. I actually am I brought an app that I'm working on for work. Um, literally internal app. You're seeing it for the first time. Um, I'm a little bit scared to show my own sort of dirty laundry, but uh hopefully I cleaned it up and I I I demoed it just now and it kind of works. So, this is a real app. This is the part. It's a real working app for my team. Um, and we Yeah, we we're actually using it. Uh, so I'm responsible for all the issues. That's why I have to be very careful. Um, so what's on screen? I You can see what's on screen. Um, I run a large u engineering conference in San Francisco. Um this uh this is an example of last year's page with 3,000 people um and tons of speakers and tons of sponsors and I I thought it might be relevant because we're at a conference like just to see have you think about the logistics of organizing something like this. Um for us we have 10 parallel tracks and so we have all these uh presentations of like well here's here's where all the the the sort of um events and the content going on that you guys might want to sign up for. Um but I don't really like the presentation. This is using an external vendor and it just kind of is ugly and like not customizable. Uh for my most recent Europe conference, I actually ripped out the front end layer and started presenting the the front end of this app which is much more pretty hopefully um that you can see here. Uh but the back end was still very much uh on a SAS software that we pay for and like don't really like. Um so for this year we are uh organizing a much larger conference 6,000 people um with a lot more speakers and sponsors and what have you. It's just doubling every single year. Um so I ended up making my own scheduling app and this is uh the AIE bot is what what we call it. So we have all these it's basically sort of a binacking problem where we have all these open slots. We have all these session days. This is a clean data set so you're not seeing anything proprietary right now. Um and we have to slot people in days. Uh and there's uh it's a very high precision task because basically if we get it wrong, people are going to show up in the wrong day. People are going to be upset, sponsors and speakers and attendees. Uh we have to sort of coordinate all of them to be to be exactly uh a fit. So uh one thing I one thing I really emphasized was uh I'm just going to try to uh demo it, I guess. Um, let's say, uh, put Scott, um, Hans Silman, ah, God, Scott Hanselman into the memory track. Um, and, uh, I I really like the the idea of like an agentic um, sort of content editor, like a content CMS thing. Um, and so it's going to give minimal detail, but post uh, and propose changes. And I really like this idea of like precision that um you know that's that's exactly the track that I intended. Um there's a I really like this idea of precision in agents because agents can hallucinate but if you have humans in the loop I think um it's really like the nice mix of UX and precision. So this is something I've been working on. Uh I'm I'm going to put uh I'm going to put a more open-ended question. Um, but I'm uh, you know, I think AIE bot is just kind of like an agent embedded inside of a web app that you can make. And I think that's cool, but uh, what's even cooler is if it lives within the communication platforms that I already have. So, uh, I'm going to go over here where I've hopefully cleared all my data and say AIBOT um, add Mark Rinovich, I don't know how to spell your last name. I'm so sorry. That's right. To an appropriate uh, track. Um, and uh, it's going to use LMS. It's going to use web search. It's going to read from my existing Slack. Um, and to me, that's how I work with my team, right? Like, we have so much coordination flying back and forth that I really want to say like, okay, um, you know, he he they've the bots's already looked him up. Uh, he's going to find like obviously he's he's a Microsoft, uh, SVP, so I'm going to put him in in that track. I can approve or reject or modify it from there. Um, and I think for me, the last thing is like I really want agents in every single surface I operate, including email. So, I'm going to approve this and go on over there. I'm going to put my friend Cassidy. uh, so we're doing it in the main application. Now, you're doing it via email. It's everywhere that you want to be everywhere, right? But because I want like your agent should be omniresent. Uh, you should just be able I should be able to text it. I just haven't set up the iMessage integrations, but like put Cassidy uh Williams. I'm gonna put Cassie Cassie Doo and see if they understand uh what her Cassidy Doo of course is her social media name. Yeah. In um I don't know. Let's let's call it the generative media track. Um uh and that's going to go over to here and show up. Um so I think a fundamental principle I really like to see is just logging and verification across apps. Um, it's it's going to take a little while, but I I sent that email. Um, and it's going to reply to me with a confirmation of uh what it thinks and what it's about to do. I can't really I I can I can reply and confirm it here or I can also see that email I just sent about Cassidy. Uh, and it's going to give me the option to uh reply once it's once it's figured out where to put her. One of the things that I think is worth pointing out is that you're using a mini model here. You're not like we don't need to use big models, Frontier models for everything. You're using a mini model. Yeah. Yeah. Yeah. I want it to be fast to be cheap. Um I can choose to upgrade the model if I want to. Uh I just haven't done so. Uh but yeah, we I have a model selector thing. So what's the harness or platform that you built this agent on top of? Uh the the the the hosting platform the cloud platform. Uh this is a Cloudflare app. Um so u I like Cloudflare at least for this V coding stuff because it can host the website. uh which is this is I mean uh and do you have an agentic harness? Are you using PI or CPilot SDK? What's your what's your loop? What's your agentic harness? Oh. Oh. Uh there's no harness. It's custom coded. It's it's a it's a while loop for LM. Uh you don't really need anything more than that. Did you vibe code the Y loop? Did I what? Vibe code the while loop around the LM. I mean, I don't think you have to. You know, you can just like do a loop and like I I like to set minimal turns. Uh I really think that uh full unbounded autonomy is a bad idea. Uh and so I like to like to have that degree of control. So uh no I I I don't think I would trust the Valu. This is a really good looking app. Like this is very clever. I can see your face like you're like this is cool. So how long it I'm guessing you spent a lot of time tweaking and tuning this thing. Yeah. And and I'm busy running my my the running the main show while I'm building software for the for the show. Uh it took about a week and a half I would say. Uh just going back and forth. Uh really you know I I think what this is it works out. So this is you're seeing like the grid view is what I call it. Uh days and tracks and all these. I also have the flat list uh which is really good for reconciling on a spreadsheet. Um I also have the raw data where I can sort of look into the exact tables that I have in uh the the sort of D1 database that I have. Um, and I have an audit log with roll back. Uh, in case I make a mistake, I can roll back uh what I what I want to uh what I want to do. So, um, this is one of those things where like I just need a ton of precision just in case I screw something up because um, you don't want to mess up the locations and assignments of 500 people. So, it's interesting because you decided to create a custom version of a scheduler instead of, you know, moved away from the SAS product that you were using before. How specialized is this to your needs versus something that other people would also find useful? And then what's the trade-off in cost of something that's not so bespoke versus your cost of implementing this and maintaining it? Yeah. Um so the the there there are publicly like private equity owned versions of this app um that cost up to $200,000 a year. Um, and uh, we are starting. This doesn't do all that functionality, but it does what we need and it's very custom to us. Like it only knows the track days of uh, June 29th, June 30th, July 1st, and July 2nd because that's all I need needed to do. Uh, so no, it's not meant for other people. I could generalize it, but why would I, right? Like those that's just time spent on things that it's not a business I want to be in, but it's just a productivity thing for my team. That's a tough thing. You make a product for yourself, for your company, it is perfect for you. Is it perfect for N plus1? Because I'm looking at this and I'm like, oh, Sessionize and all these folks, they should they should buy this from you, but you're like, no, it's fine. It's just for me. You just made it for yourself. Yeah. Sessionize is is uh one of the industry standards, but they don't have a lot of the the things that that I would want uh to upgrade to. It's like a C vent or an Excel form. And I can imagine add talk, add voice to this. You could just go for a walk and talk to the agent and it makes the moves. I happen to have an opinion that I think people adding voice into their apps are doing like the locally optimal thing, but it's not globally optimal. Globally optimal is your voice should be tied to your OS kind of so that you have one hotkey. You don't have to learn like five different hotkeys. Right now, if I press command K versus command P, I have like three different apps all trying to be my voice layer, but that's like really messy and like it's not going to really have the memory of like what I want to say, not not learn on my uh my my patterns. Uh so I actually have this like I do I do like having that separation of voice versus the rest of the You're using window key T. Yeah, I use Windows H or Windows key. Yeah, but I agree that too many people are wasting time adding voice into their agent apps. The OS should handle that. It should be available for that. Last question as we get towards the end here. Um, is this on GitHub and use GitHub actions? How does it deploy? Very briefly, how do you deploy the app? Yeah. Um, uh, Calfare has a a sort of wrangler CLI that can that can push everything including migrations of a database. Uh, and it's all done in GitHub actions, of course. I got one one more question. How much of the code did you write versus AI, right? Uh, is this like the the sort of reveal moment? Yeah, it's entirely viodated. What? It's entirely viodated. I I I yeah I was forced to lie about the the loop thing. Uh even that was done by Gemini. Did you review the code at least or did you just play test it? I I just I just tested and give it feedback. Um so you can see um all the all the feedback that I have in cursor. I had to hide this window um where this is the entire history of all the vibe coding that we've that we've been doing. You never looked at the code or really worried about the code. Uh no, I don't know where I don't know where it is. It's vibes. Yeah, I I will say um I do I do care a lot about parallelism of agents and so if you let an agent just run, it's going to do one monster file of like thousands and thousands of lines. Uh and that's where I draw the line. I I have some skills where uh it'll keep the the code maintainable but also uh parallelizable so I can have like five different agents running on it at the same time. Entirely vibe coded. Absolutely fantastic. Big applause. Best fives. We're going to go to our sponsors. We'll be right back. I got one more. Microsoft Build in San Francisco and I'm very excited to be here in the Fireworks AI booth. I'm here with VC and we're going to tell you all about this amazing technology. VC, welcome. Thanks for having me. Absolutely. So tell us a little bit about what Fireworks AI actually is. Yeah. So, Firebox AI is the production AI platform for open weight models. We process more than 30 trillion tokens per day. We serve more than 10,000 enterprise customers globally including some of the biggest names such as Cursel uh Versel, Uber uh on the platform. And these companies leverage this not just for world-class inference but also for fine-tuning the model on the platform so that they can truly embed the data mode into the platform. cursors, Composer 2, Composer 2.5, leveraged our RL platform. Um, and yeah, this this is the this is the thing they need to actually scale up their applications. Oh, that is so interesting. So many new things being invented. It's a little hard to keep up and I'm just loving so much that I'm learning here. What is important about the partnership that exists between Microsoft and Fireworks AI? Yeah, so absolutely. So, we are already providing worldclass inference to customers. Now, they can access it within Azure Foundry. So they have access to the Azure foundaries governance, security, reliability, no need to worry about billing and stuff but they can also get access to worldclass inference which we provide on the platform. Um and then also we are bringing much wider array of models into the platform, openweight models into the platform and they can train a model on fireworks and sub inference on Azure. All those meaningful features and simplified billing. You had me there. That's that's fantastic. Uh and so I know that this is your first Microsoft build, right? Yes. Are you excited to be here? Super excited. I'm excited to meet people at the booth. I'm excited about the uh talk on Veterans Day which I'll be doing along with my colleagues who are giving some talks. That'll be a lot of fun. That is great. Listen, you can find demos of Fireworks AI on their sponsor showcase page. Uh plus you can find links to the sessions that are being given so everyone in the digital audience can follow along. So, thanks for being with us and we'll see you again soon. goals. Build and deploy Genai applications with Elastic using Microsoft Foundaries catalog of LLMs, semantic kernel, and AI agents to improve search relevance. Enable scalable AI from cloud to PC and optimize Aentic AI apps across Azure Cobalt and Windows on ARM. Find out more about how our sponsors are using AI to solve the issues of today and accelerate the innovation of tomorrow. All you have to do is visit our sponsor directory for Sponsors make more possible. Hey friends, we are back with Mark and Scott learn to vibe check. We have had some amazing stuff from Swix, from Steve Sanderson, from Cassidy Williams. And now to bring it all home is the legend himself, Simon Willis. You may have seen him occupying the top spots on hacker news on any given Tuesday, pushing Mark and I to the bottom of the second page. Simon is a legend. You can check him out uh online. And today, what have you brought us, Simon? So what I've got today it's the culmination of eight years of all of my projects finally come together. Um the first project uh is most am I shown on screen somewhere? You're seeing your screens right now. Fantastic. So I've been building this project called data set for like eight years. It's a data exploration tool. Lots of interesting data exists in the world. What can we build to help us explore it? Three years ago I started work on LLM which is a command line tool and Python library for talking to language models. It's got plugins for basically all of them. And I've been blogging for it turns out 24 years. So I've got 24 years of stuff on my blog. So obviously the thing I've been building is an agent because everyone's building agents like the the hello world of programming in 2026 is to build an agent. This thing is called data set agent. It's a plug-in for data set and it lets me ask questions about my data. So let's do count of entries and quotes and link and I call them blog marks on my blog. So I'm kicking this off. This is using GP 5.5 under the hood at the moment, but they'll work with any of them. It has a bit of a think. It looks at the databases and queries that are available databases and tables. Runs a bunch of queries. It goes, "Ooh, some of them are draft posts. I have to think about that." There we go. It says I've got 3,000 entries, 1,000 quotes, 8,000 blog marks. I'm going to say do a chart. So data set agent as with everything in data set is based around plugins. One of the plugins I built is called data set agent charts and that can do little bar charts of things. I've got other plugins for running bits of Python code in a sandbox and there are all sorts of visualizations and things that I want to add into this in the future. I will do one more demo because I've got an that one was GP 5.5. This one here is the same software running against Quen 3.5 on my Mac in LM Studio. So, let's let's see if it does the same thing. Um, count count the entries and let's just do count count the entries and quotes. So, now this is running entirely offline. This is using Oh, that was quite quick. Very nice. Yeah, it knew the tables. It did the counting. This worked. Quen, these models that run on laptop are very capable of writing SQL queries and doing basic tool calls these days. Okay. And the tool itself, the tool here that we're watching running on local host is speaking to it has tools available to all of the things that you have on your blog or is it how did it know which where the data came from here? So it is running a copy of my blog's database. My blog is Postgress on Heroku. I export that into I built another tool called DB2SQLite that exports that to SQLite. So I've got a 100 megabyte SQLite database file on my computer which data set is looking at. And so all it gets is a SQLite database and then it has to read the tables and get the database metadata and all of that kind of stuff and make that available to the agents. So first Simon, I'm a huge fan. I love the Pelicans on bicycle. Uh I've got a demo about that. Okay, but I got a question about this. Um, when you say plugins, what is what is a plug-in? So, this is um I built everything in Python. Python is a very good language for plugins because it's all dynamic. You can install extra packages that become visible to each other. So, a plug-in is actually let's pull one up. I'll tell you what, I'll pull one up that I wrote an hour and a half ago. This is a data set agent microython. It is live on the Python package index as of 2 hours ago. And this is a plug-in for data set agent that adds the ability for it to run Python code in a web assembly sandbox because I want agents to write Python code. I don't want to delete everything on my computer. You need some kind of safe way of doing that. And if you look in the code, um, it is and this is what you were working on when I saw you in the hallway because I can't talk to you. I'm publishing this this thing now to to Python. This is exactly what's going on here. Okay. And make the fonts a little bit bigger there for us, brother. Here we go. So, a plugin just has a decorator called hook impl because it it implements a hook register agent tools and it returns a list of agent tools. In this case, it's called execute microython. It's got a function. It's got a description. Um, and all of data set itself has well over 200 plugins now. LLM's got about 40. My little agent thing has got, I think, five in the in the wild already. Plugins are such a great way of building software in this age of AI because honestly claude code and copilot and things they can oneshot a plugin for some quite sophistic behavior if you give them the right kind of information that starts. So plugin instead of a tool. Well, the plugin is a bundle of tools. Like in this case, I've got one tool that's embedded in a plugin. A plug tool is an MCP tool. Not MCP, but the same idea. MCP is just a fancy wrapper around the underlying concept. But you you think this model works better. It I mean it's because you're working entirely in Python, right? Exactly. Exactly. Like MCP is just tools with extra steps. Now, when you do this kind of work, if you recall, folks need to remember that he's had a blog since 2002. Like, we were sending back trackbacks and ping backs on uh blogs 20 plus years ago, Simon and I. That wasn't always vibes. When did you start vibe coding? Because so you've put so much out that is bespoke code. How are you incorporating the vibe coding and then the AI augmented software engineering practices into the work that you're doing? So about two and a half years ago, Claude's artifacts came out and got really good at building little HTML and JavaScript tools. And so I started vibe coding before anyone was calling it vibe coding where I just do little self-contained tell self-contained HTML apps and I put those on tools. Simon wilson. This is a collection of 215 of these things that I've been building up over the past couple of years. And this is great because it's so safe. like you can't shoot yourself in a foot in the foot with a static HTML and JavaScript thing on separate domain. So there's there's sort of no risk in in these being mistakes. And I'm still doing this today. Like um I built this one yesterday. This is just a prototype of what it would look like to build an interface where you can attach files to a text area and like preview them in line like this and copy and paste things in all of that sort of stuff. I want to build this as a feature in my agent. But first I built the UI prototype and this I built this with um this one was claude codeex was was um codeex desktop. Often I use clawed code on the web on my phone. So I can fire up my phone and say hey build me a prototype that does apps. Click a button on my phone to push it to GitHub. It gets deployed by GitHub pages. It's live. So I've been building little throwaway things the past couple of years. Since January, most of the code that I write has been written by an agent for me. Cuz that was the point when the coding agents got good enough that you can pretty much tell them the style that you like, how you like your unit test to look, all of that kind of stuff, and they do the the sort of scut work of actually typing it into a computer for you. So, when you're uh vibe coding, do you watch the thinking trace to see if it's off the rails or It depends on the stakes. So low stakes things like a little UI prototype, I don't care. The code doesn't matter. I just want to prove that the thing is possible. Some like the um this thing right here is a sandboxing system that runs untrusted code in a sandbox within my system. That's really important. Like the stakes could not be higher. This one I watched it like a hawk. I was paying very close attention to the code that it was writing. I then ran all sorts of tricks to try and um like stress test it. And I actually had GPT 5.5 try and break out of the sandbox which was quite entertaining to watch. Um found a couple of edge cases that I needed to clean up. But yeah, that's um that's the sort of other other end of the scale from just hands off letting it do whatever it wants to. So one of the techniques that Scott really likes is adversarial vibe coding where you have one model propose something or write something and then have another model go critique it and and then have them argue. I've been doing that a bit. GPT 5.5 is my security blanket. I really like it for security reviews. So, until recently, I was doing most of my work in Opus 4.6 and then I'd have GPT 5.5 do a quick scan at the end to see if it missed anything. And it would often find things. To be honest, though, if I'd asked Opus to scan its own work, maybe it would have found them, too. It's just you sort of habitually feel like you get the sort of vibes of which models have certain strengths. Yeah, I like to have two or even three and I make a high, medium, and low. And if all three models agree that something's a problem, then that's a high. Two, it's a medium, and one, it's a low. So, being able to use something like GitHub Copilot and the CLI or the app, you can pick two or three models and have them all fight and it works great. Yeah. So, are you no longer typing code? Are you are you a typer of pros? Do you type your prompts or are you a yapper? Um, I'm still mostly typing the prompts. If I'm walking the dog, I might dictate prompts into my phone, but even then the prompts can be quite short. Like something I found really interesting is it used to be the harder the task, the harder you had to concentrate. The opposite is now true because the task is really hard, GPD 5.5 will turn away for 10 minutes and during those 10 minutes, you can go and read hacken news and you can get distracted and work on other things. That's this weird thing where the more difficult tasks are now cheaper to be distracted from than the easy ones are. So this is the thing that Mark and I are struggling with and I think what we'll talk about in our close is this idea that you have to know what you're doing to be successful at vibe coding. You don't vibe into production. There's AI augmented software engineering baked into your process by virtue of how your brain works because you've been doing this for 20 plus years. Would you agree? I think AI has made being a professional software developer harder because it's raised expectations for what we can do and you can now take on so much more ambitious projects. It used to be we had to stay on our lane. Like I was a back-end Python programmer with a little bit of HTML and JavaScript because I could only skill up to professional levels on one of those technologies at a time. Now I'm writing. Um, here's a little app I wrote in Swift UI that helps see helps me see what bandwidth my computer is using. And this is another one that shows me what's using my memory and what's using my GPU. I don't know Swift UI. I have no idea what I'm doing. I didn't even have to open Xcode for this thing. Claude Code is like, "Hey, no Xcode needed. I'll just build this up and stick it in the menu bar for them." I would not recommend other people use use or trust this piece of software because I don't know what I'm doing. But it's been it's it's also I've been using this software for like four months now. I kind of trust it now. Like I've realized that the thing I care most about isn't that an expert wrote the code. I care that a real human being has actually been using it for longer than like 2 hours and it's sort of built up that like over time that trust in the software gets built up. Fantastic. Well, this one feels like this one feels like AI augmented I agree with that. One closing demo just 30 seconds before I'm going to kick you off the stage. When did Simon last see a pelican because it's got all of my wildlife cycle now. You saw a pelican last night. And it says, "I saw a pelican on June the second today." And it's right because if you refresh my blog, there's a pelican. They're right outside diving into the water. It's super cool. Very, very cool stuff. Big congratulations to you, sir. Thank you very much. That is Thank you. Thank you very much. We appreciate you. All right, we're going to say a quick thank you to our sponsors and we thank you all for hanging out and then there'll be a brief ending uh here on Scott and Mark Learn to Vibe Check. Welcome back to the floor at Microsoft Build in San Francisco. I could not be more excited to be at the Qualcomm space. I'm happy to be with you because we've had so much that has happened in the keynote together and you're going to tell us more about that, aren't you? But first, tell everyone who you are. Hi, I'm Anna Schaefer. I'm responsible for Qualcomm's developer and partner program and talk about partnership, amazing partnership that our two companies have had throughout the year. Uh, exciting things that we talked about in the keynote together. It is really fantastic. Now, what does this community need to understand about the types of innovations that you're really delivering and how we're doing it together? So, this is the age of AI agents and you know, think of these agents that are contextaware that are taking data from across different devices, you know, from a ring to a pendant all the way up to the data center. This means very new ways that developers can deliver experiences for their their customers as well as different ways that they need to work. It sounds like that's a lot of new workflows and form factors to understand. How are people going to actually dig in with this and learn? Yeah. Well, thankfully that partnership with Microsoft includes very tight integration with things like Foundaries Factory, Windows ML. uh we in our sessions and our demos in the booth. Uh we're going to be talking about AI agents that uh help you port your apps native. We're going to be talking about automation in your workflows for inference so you save on those tokens and think of that orchestration that is needed as you make those workloads go across these different form factors. That is fantastic. Well, I know we're really excited about this and you know how you can be even more excited. Go check out the sponsor showcase page, demos, links to their website, all of the new information that's available. That's how you can stay up to date on everything that's happening between Microsoft and Qualcomm. Thank you so much and we'll see all of you again soon. And we are back closing up. Mark and Scott learn to vibe check. We have seen an amazing lineup of Were you blown away? Yeah. So, you didn't know this was going to happen? No, I didn't. I didn't. And and those were each one was amazing and and unique. In its uh own way of what how kind of it showed what you could do with vibes/ monitoring and kind of more on the engineering side of things, right? We started this show saying that we were going to see slop. We were going to see vibes. We were going to say AI augmented software engineering. I did not know. You did not know. None of us saw any of these projects before. To be clear, I think there was one slop. Which one is that? the CSS. I don't know, man. I thought that was genius. I got to shout out to Cassie for that. But I thought that the Vibe OS was amazing. All of these are It was just so pointless and wrong, though. Yeah. But I think the thing that we can both agree on is that you have to know what you're doing and that you will always be more successful if you try to apply good software engineering practices when your apps to to have those ideas like Cassidy saying that she couldn't have done that without knowing CSS. I don't know CSS. I don't think I could have talked my way into that. Yeah, that was amazing. I want to say shout out to everybody who made this possible. Shout out to our sponsors. Big thanks to Steve Sanderson, Cassidy Williams, to Swix, and to Simon Willis. This has been Mark and Scott Learn to Vibe Check. I want to encourage you to check out our podcasts. We do it every week. Mark and Scott Learn to where we learn a new thing every week. I hope you have a lot of fun at Build. And thank you for spending time with us. We'll see you soon.