I Built 52 Startups in 52 Weeks with AI Coding Agents - Here's What Actually Works

0:00incoming transmission from the future Vancouver AI community meetup >> and I'd like to invite you to stay for the last 20-minute talk about AI coding agents from Michael Yagudaev. Michael, you still here? >> Yeah. >> Nice. Um, Michael built 52 companies last year in 52 weeks and killed 51 of them. He knows more about AI agent coding, agentic coding. He's changed his workflow four times in four years. Um, this guy's our local source. Um, there's this whole idea of building communities in or it's not building communities, I'm sorry, building companies in public, and you should look it up. It's a movement.

0:43It sounds simple, but when you talk about building companies in public, this guy is walking the walk. So, they share their revenue receipts online every day to show the growth. They post all bugs, all response. Everything that they do happens in public online and he's done a great job of modeling that modeling that. Um, I am attracted to that way of operating and and try to do it, but he's committed to it and I respect that about him. Michael, come hang out. I didn't draw an X in the floor, but I want you to hang out right here for YouTube's sake.

1:16>> Okay. >> Thank you. Right. Right here. So, uh, before I do that though, um, you all been sitting there for a long time. I want to invite you to just get up for a second and stretch cuz, um, yeah, sitting is killing us all. Just take a minute, stretch, curl, do, uh, do this.

1:38Yeah. Thank you. Cool. Thank you guys. Cool. So, um, is it working? >> Okay, perfect. So, you want me to stand right here, Chris? >> All right, sweet. So, uh, Chris kind of introduced me a little bit. I want to talk about AI coding and um, I built 52, well, I tried to build 52 startups uh, last year. So, I was sold a previous company that I had and uh, I just was trying to figure out what the hell to do next. So, I thought um you know 12 startups in 12 months isn't cool, which is what Peter Levelvels talks about. I thought 52 startups in 52 weeks is cool.

2:24Uh this year you should guys probably do 365 startups in 365 days. Um cuz why not? Um but in all seriousness, I did that just to figure out what to work on. So, it took me uh 12 startups and 16 weeks to figure out what to work on and then I can just stop because there's no more learning there. Uh figured out the startup that worked for me which is audio wave. Audio lets you listen to text as high quality audio. Uh built it for myself, went viral. Kind of kept building it and then went to um this year I was kind of bored so went to an SF startup and I was trying to build a

3:04second brain AI. Um, so work with our beautiful American buddies there trying to do that for a few months and then um, Claude Code came out and I had a bunch of new ideas. So I'm like, hey, what the hell not try to build all of those ideas, right? And then uh, I started thinking was I faster?

3:27Does it matter? And surprisingly preparing for this talk, I thought it was a yes. And then I presented to Chris and I realized it was a big fat no. And why is that? Why is that? So I had to think about other people and maybe it's just me. Like you probably felt that way. I feel that way for sure. Um why is it that we don't have five new apples starting today? Like why don't we have high quality startups um there? So, cuz we're delivering more output, but we're not delivering more outcomes and we're not delivering more polished apps.

4:12So, we can build more stuff, but we're not building anything that people uh actually care about. So, customers don't care if you have five people that you build your startup with. Uh your Spotify is built by thousands of people. Do you care? No. You want to stream a song and you want to listen to it without interruption. And that's what we should be building stuff that people care about. Not less it's going to cost us less money. We should just figure it out. That's the job of me as a founder is to figure it out. It's not your job and I shouldn't be talking about that with you. Uh I should be talking with you about keeping it simple and

4:53delivering something beautiful. So two concepts here that I want to introduce is um so back in 1995 Paul Graham uh actually was able to build what is today kind of his version of Shopify was the first web app out there called Via web via web uh e-commerce platform if you don't know Shopify I don't know what you're doing here but um yeah so he built that in 1995 first web app there and he was talking in one of his essays is about I was reading one of his essays. He was talking about that he they're able to ship things so fast that they can be on a call with a customer and then iterate and fix a bug for a

5:37customer on the call. When is the last time anybody has done that? I've never done that honestly. Maybe once, maybe twice. Um and then I started thinking maybe it's not about tech. Maybe it's about keeping it simple. And that's what allowed me to build all these 12 startups in 16 weeks was keeping it simple. And then the other thing that uh enables us to build things quickly is this thing called leverage. So leverage is all about increasing our output with the same input. So in here we have a lever. Uh a lever is just a tool that we use to lift a bigger and bigger rock.

6:19And the further we stand on this lever or the further we uh push on this lever here, the more force will apply to this rock. But we can also reduce the size of the rock that we want in there. So that's reducing the scope, reducing the friction. Uh and that's what that's all about. So it's a balance between these two things. Um and how do we get a balance?

6:45Well, we need to start with something really simple and then we extend leverage as we build more tools. We can't over complicate our tools and that's what kind of started happening with our tools. We started with uh really simple tools. We can build stuff. So here uh one one really easy way to do this is to try to build the simplest dumbest thing you think you can think about and then see if it works. So here a customer asked me at one point, what languages do you support? And even though I support many languages, uh I only focus on English because that's the language I speak. That's the language I listen to content in. Um but I wasn't

7:27showing that all these voices can actually work with other languages. So it was really easy to just add that in there, right? But it took four and a half hours to do that even with AI. And it took that long because I needed to figure out what to do. I didn't have a design. Still had to go through all these motions. And it was still uh I did that by just breaking down the tasks into smaller tasks and thought about the simplest screen I can put this in. This was the simplest screen I can think about. From a UX perspective, didn't have a lot of constraints. So the the smaller space you have, the more you

8:07have to think about your real estate. So in here I have a lot of space. I can do a lot of things. So it was nice to build it build it with cloud code. Um I'll talk a lot about cloud code. Don't worry. A lot a lot about cloud code. I'll give you my referral link so you can give me more credits on cloud code at the end. Uh just joking. But um there's a simpler solution to this problem. I realized after I was writing article about it and I realized, oh my god, I could have just put it in the damn FAQ. What the [ __ ] I could have spent 5 minutes, not 4 and 1/2 hours,

8:43and solved problems for a customer. So, it takes some training, and I feel like sometimes you kind of kind of like going to the gym, you lose it sometimes if you don't use it. And then um cloud code is actually another really good example of not being fancy and not using all these fancy tools. So when they implemented search in cloud code instead of using like the cool state-of-the-art rag which is retrieval augmentation you probably heard about it but you have to maintain a separate data store and kind of sync between the chunks of code that you have index them uh and maintain all that. So the cloud team decided what if we didn't do that? What if we use the model to

9:26just call really simple Unix tools like find and grip and so on and search through and see what happens. So they did that and guess what it was just good enough and we still use that. So um sometimes you just don't need to go through the complexity. You can always do it later. Uh a lot of times you will as a startup you end up in this complicated mess because you were originally doing something a lot bigger. You kind of want to conquer the world, right? You wanted to work on every platform everywhere all the time, but you end up working nowhere really well. So I was working at the SF startup. We were building uh originally

10:09it was it's an SF startup, right? So we have to think big uh which is a good thing. So we're trying to build an OS that will actually give you context on everything that's happening uh in your life and give the LLM context. So every every everything um we start by like forking pretty much Linux I think but then it was like okay let's do it to Android and then it was like oh we can build a Mac app. Uh so we build a Mac app there. Uh, but it's like what about Windows? We want to build Windows. Um, we don't need to build Windows. We just need to build a damn Mac app and we need

10:46to integrate really deeply into it. And we have four freaking programming languages there. No amount of AI could help us with that. I promise you that. We try every day. I try every different model that uh was released every time and none of them helped. So, um, yeah. And then I just simplified it to uh the team actually simplified it to just be uh Swift UI and they still have Python.

11:11My next desktop app was just Swift UI and TypeScript and I used my basic knowledge of Swift UI with cloud code to build everything and organize things. So that was kind of the power that cloud code gave me. So as I was saying the tools starting to get really complicated and you kind of see it we extend the lever over and over and over again. It's again it's a good thing but we have to kind of take it one step at a time. So first we had GitHub copilot. GitHub copilot just let us drop a comment in a file and then like something like a calculate add a function to calculate taxes. He'll calculate the sales tax for you. Uh

11:55boom, you're done. Awesome. um very limited. You still have control of your code. Next came cha GPT. Chiao GPT you can actually write ask it hey write me a script to calculate tax on all of my products or my customers invoices whatever you want. Uh and copy and paste that drop into your code and you have it kind of working. Um then we integrate it into the environment where we code. It's our IDE. We integrated into cursor. We integrated into we created a new IDE actually called cursor and windsurf. Uh and we can actually start having conversation inside of our code editor.

12:34And then we can take it a step further and we started building a little agentic flow in there. And the simplest agentic flow in there was just um writing the code and then running it through a compiler or a parser on your system with like a simple like I think it was even preempt but it would run these tools automatically and then if it um didn't compile it correctly it would just um do another loop on there as well and claude code came out and then we had to throw out the whole freaking thing and start over uh have an identity crisis and like I don't know if any of you had that happen uh had to change my whole

13:15workflow. So with cloud code instead of starting by looking uh instead of asking it inside of my editor and seeing all my files open I would start with the task. I'm like holy [ __ ] I'm a project manager now. What the [ __ ] do I want to be a project manager? I mean that's okay but um it worked really well. Um and I'll show you my workflow with claude code there uh and the limitations.

13:43So the other thing I want to talk about is this Andrew Karpati's uh version of software 3.0. So 1.0 was a software that we used to write before with Python. So uh if this then that happens kind of very rigid instruction, very predictable, very nice. Uh and most of software you know is written like that. Then we came up with um machine learning. It's like, hey, here's a bunch of data. Um like do some teeth whitening for me. So here's a before and after of teeth whitening and let's actually do the teeth whitening there. So we have a neural net and we figure it out in the fusion model. Um and then the generative AI came out and generative AI was more

14:24of a general. So with 2.0 we just focus on a task. With 3.0 we're general and we start having um LLMs that can do anything we want. create essay, summarize things, uh, and so on. And, uh, we can actually now, interestingly, go back to 2.0 or 1.0. So, our 3.0 code actually ends up generating so much code in the 1.0. So much so that we can't even review it all. Like, I saw 125,000 line pull request on an open source project. Uh, what the [ __ ] Nobody can review that. So um so there's paths we can go and we can go from 3.0 to 2.0 by actually prompting it correctly. So we can specialize it by saying something

15:13like hey um only change the person's smile or detect the person's smile. We can also ask it to uh write a script to detect the contours of the face and it's known algorithms to detect stuff like that and only do an infill on on that. So we can start with 3.0 and kind of move it there and figure things out. Uh, and that's how we build stuff. Now, so that's that's leverage there. There a couple of really good examples to inspire you here. So 3.0. So this is uh v 0ero and this was v 0 of v 0 uh when they first started. The amazing thing they did really well was that they just put a uh text area here. You put your

15:56text into it. um you put like UI tweet and a couple of examples on the bottom and at first it didn't even work. What it actually did was because I cuz I did that I was like, "Oh my god, this is so exciting. All my UX can be designed now." So I s I tried to do that and then it's like, "Hey, sign up." Like, "Okay, cool. Signed up. Oh, we'll tell you when it's ready."

16:20Like, "Okay, what the fuck?" Um but that's okay cuz they validated that we really needed it and then they can test with a smaller group and that's what you can do as well. So any of us can kind of do that and that that was the brilliant thing that they did there. Um and they also used the um the basic just create a basic wrapper around the model. So the next one is this remove BG. It's the 2.0 just kind of what it sounds like. It's pretty self-explanatory. removes the background like train it on a bunch of images. It was way before all of the open source models that we have now. So, uh really really good use tight scope.

17:02And then uh the dumbest one is this uh which probably makes a ton of money. Actually, I haven't looked at how much money you can open your chat and use all that energy and find out for me. But, um it's a really simple and actually brilliant idea if you really think about it. What it allows you to do is it allows you to create a website of just links and kind of a social media profile kind of just a couple details about yourself not a WordPress blog. You don't have to configure anything, buy anything. It's really really easy and simple. Um then just brilliantly done.

17:36So that's that's beautiful software there. Couple of ideas to that I use to get more leverage on uh my code. So um you can test a bunch of ideas in parallel. So talking about overusing, I'll show you what it means like after. But yes, I do run like four or five cloud code sessions at once, sometimes more. My friends run 10 or so in parallel and um I just switch between them. I give them a task, switch between them and um ship it as as much as I can.

18:11So the other thing you can do is um let me see what I wrote there. Um oh yeah we can ship just more of ship the boring stuff right ship the health documents ship the marketing material ship the metric there that measures a feature itself. Uh and then we can uh see what's going on and kind of uh we can also ship a variant like a four variant like when you first use image models you probably saw the four grid square of different things. So we can give CL like four different cloud code instances create a login UI and see what it comes up with.

18:50We can also be more explicit and tell it uh each one of them to do a different thing because you want to explore uh a complicated problem and you don't know what the best approaches there. So that's really useful to to innovate really fast. Um we can also just ship things just uh run it overnight a lot.

19:10And uh what else did I have there? Um yeah, that's pretty much all I had to say there. Oh, and we can be uh datadriven uh actually here. So, we can actually feed a post hog MCP into it and ask it questions about and and have it run experiments automatically. Here's an example of somebody really really talented. So, he's a AAA game designer that built this echo of somewhere and use a lot of AI to create like in game development the creation of assets is really really timeconuming. Uh the game mechanics is also hard but you you just have to generate a lot more assets than we do in our uh when we build apps. We don't need to generate as

19:54much assets. So he use it really well. Uh just beautiful game and he still uses taste and leveraged his skill set and um if you don't have the skill set you build this and still make a ton of money. So Peter levels built this game here. Uh, what if I told you this game generated over 100K in revenue to Peter and he just vibe coded it and he had fun. So, um, just a masterclass of business there, but just a little bit maybe you should learn from that first guy there, but that's okay. You have to start somewhere, right? Uh, and if if you're going to build anything like a vibe coding game over a weekend, it's

20:39really really fun. So, uh, I guess the party can't go on forever, unfortunately, uh, as I found out this week. So, we have, um, 29 days. I think it was July 31st, so this is August 28th. Yeah. Um, so you have 20 29 days to get as much co cloud code usage as possible. That's what I read in this email. Uh, and then replace an open source tool. and Kush there will will give us a talk about it hopefully or Chris will find somebody to talk about Kimmy K2 and Quan Coder. Uh I need to know all about this in a in about a month when I run out of all these credits that I'm paying $200 a month

21:20for. So um all joking aside um I do find Claude Opus to be really really good at complicated tasks and I don't know I just maybe I just have a lot of complicated tasks. Maybe my life is really complicated. Um, I don't love it when it drops the clouds on it. Generally, it just it generates more bugs than it fixes. So, um, and that's why I paid for the, uh, $200 a month plan. We're constrained by what we can do. Obviously, like AI companies like to tell you, it's like you can do anything, but uh, we only have so much time on this planet and we can only do so many things. And the best

22:05thing to constraint I find is time. You can never buy more time. So put a constraint on time and see what you can do with extra leverage. Keep it simple like reduce the scope. That's one way. And go get some money from the Americans cuz they have a lot of it and they should spread it around a little bit more.

22:27U build big things. Like don't be afraid. I feel like we're afraid to do that as Canadians. So another constraint we have um is this amazing study that came out in June. Uh it talks about the cognitive um pretty much chat is making you dumber. That is what it's saying here in a in the nicest way possible. So they did a they did a study on multiple groups of people. So the first group wrote so they all wrote essays SAT level essays. high school essays, guys. I high school essays. That's the level that that we tested at. Like, come on. Um, we're not talking about like PhD level stuff. So, even at high school level,

23:13uh, the people with on the brain group are just not they're um only using their head to write their essay. Uh, the people with the uh search tools, they're using Google and other tools for the traditional search. and then the LLM people like using ChachiBT in there to write their essays. And what they found is that the activation in the brain in the people that are writing with just without any use any tools like shows all these like beautiful connections. I don't know if you can see that but all these beautiful connections here. the search team just um yeah that's that's not it's not great u and they found that people are actually losing their ability to do

24:01critical thinking on the LLM side so they and it got worse so they used it for months and months and study was going for months so the first couple essays they kind of adjusted a lot of things and um still did a little bit work and thinking and had recall All of what they wrote about had some attachment to the topic, but by the third fourth essay or so, they're just copy pasting chat GPT. The evaluators described it as having soulless essays without any understanding. And after after they're actually forbidden from using chai GBT anymore, uh the worrying part is that they're not able to produce the essays like the first group which is really interesting. So it's actually a

24:48prolong effect. So it's like a muscle. You either use it or lose it and we cannot lose any more critical thinking in our society. So this is this is a warning there if you're not watching the news like people do not have critical thinking skills. I hope people in this room have it but they do not have it uh unfortunately.

25:10So um to that I said don't outsource your thinking like write your initial task yourself. Think about the initial solution how you would build it and then write it down and have the LLM help you come up with a better solution or different solution. So do some of the work yourself and you'll build better software that way too. I found Yeah, thank you.

25:41Um, so LM produce a lot of crap. Uh, one way to kind of add some quality control into it is to close the loop what we like to say. So give it a feedback. So I was mentioning earlier the first closing of the feedback loop is just compiling the software. But we can actually add automated tests into it. So we can I in my claw.m MD I have it write an automated test first run it make sure it fails then after it fails write the implementation there and it will write much better code in general it will run itself for uh AI for UI related sorry tasks uh what I have it do is I have it

26:21take a screenshot save the screenshot and then critique its own screenshot and how close it is and um that works really well and I also look at the screenshot after it's done to see that it did the right thing. The next thing we can do is we can actually create like a reviewer.

26:36We can ask it, hey, take a look at what you did. Was it good? Uh, oh, maybe not. Uh, and it'll always say, oh, absolutely. I see what you mean now. Uh, I should not write 5,000 lines of code. And then it goes and writes 5,000 lines of code anyway. Um, but it it is getting better honestly with that. So, just you have to prompt it sometimes. sometimes you have to put it on a tighter leash.

27:01The last thing I want to mention there uh with that is eval. So if you're writing 3.0 software, if you're using a lot of LLMs in your software, it's very unpredictable. Uh and as every time you change the prompt, what happens is that you have um you can actually make the system worse.

27:21Evals are a way to create a benchmark pretty much question and answer known answers to questions and you just run your questions again and see if the answers are close to uh what you're expected. You can also have an LLM as a judge that evaluates these answers like how good was the answer by another LLM.

27:40Um so and this a whole topic that Chris can probably invite somebody smarter than me to talk about. Um do you want to finish up? Will you jump ahead to your um your workflow? >> Sure, >> Michael. >> I I will take >> You're welcome to double back later if the questions come up, but this is the [ __ ] that people want to see. Bro, >> this is the magic. This is like not very many people do this and the few that do wouldn't show it to you.

28:09>> Sure. Um, so let me talk about my workflow a little bit. So, like I said, I like to write my own. And uh I start with a really simple linear ticket usually just something really quick write it down break it down into a couple of tasks and then I'll think about how to implement it probably uh hopefully hopefully I don't get lazy and get into those um third group there. Um after reading that after reading that um study from MIT I kind of I want to do more of this uh thinking now. So uh I write down what I want. Uh I just give it to the MCP server to uh pull it from linear. I run

28:51cloud code with dangerously run without permissions cuz I don't have any crypto on my machine and what's the worst that can happen. Uh I see you smiling laughing there and you probably tell me a lot of bad things but nothing has happened >> so far. Let's go. >> Um, yeah. So, uh, then I, so I have it generate a plan generally for me, sometimes a PRD, but, uh, I review the plan. Sometimes you'll add like a lot of stupid stuff in the plan that I don't need. You will try to think of all the edge cases that are just not important.

29:24So, that's one area to just squeeze away and remove a lot of the complexity. So instead of uh a login system and and doing a two-factor authentication, all sorts of crazy stuff that it might come up with, you can just delete delete delete delete uh and that give you good results generally. Then I just ask it to execute the kind of mentioned closing the loop and running the automated test.

29:48I have that in my cloud.mmd. So it'll do that automatically for me. And um and then right there, I don't know if you can see it. Uh once it's done, it will ding me. So it will actually play sound or something like that. Um and it will let me know because I have like four, five, six of them running.

30:08>> That's the longest you ever got one. >> Uh I don't know cuz I just get too bored and it feels too long. So it feels >> leaderboard using cloud codes. Report back on the longest time. >> I I I've run it overnight. So I have no idea. I just Oh, I use Caffeinate. It's a Mac app that lets you like keep your computer asleep uh awake, sorry. And uh so I'll caffeinate it and then for 4 hours or so and I'll go to sleep giving it a task and hopefully I don't get kicked out of cloud code for that. Um but who knows? So that's pretty much what I do. A couple of honorable

30:44mentions here. So Cyrus is a tool that lets you link your linear to cloud code and you can assign a linear task to cloud code and Cyrus runs on your machine. It listens to new tasks that come up. So if you assign Cyrus to cloud code uh to sorry if you assign Cyrus to linear ticket it will pull down that linear ticket and start executing it whatever is in claw.md it will start doing. So it'll start implementing the ticket. And what you can do with that is if I'm at the gym and I have a bug in my mobile app, which happens sometimes. I can just write a quick linear ticket, boom, assign it to uh using Cyrus and

31:28then when I'm back to my desk, I can take a look at it and fix that bug uh and get that out of my mind and build better software that way. Um, yeah, I use I kind of switched the cursor lately back. I went to windsurf, not cursor. Uh, the reason I switched back to cursor, hilariously enough, was that I had too many windsurf rules in there and it was starting to be too slow and I was already using cloud code anyway. Um, so switching back to cursor was really fast cuz there no rules. Probably if I just delete all my Windsor rules, it'll be better cuz I use it as a Q&A more. So

32:05I'll ask it specific questions. So claude code I'll ask it to perform tasks for cursor. I'll ask it like specific questions like what did you mean right here? Um kind of thing there. And I use a repo prompt uh and other codes. I try to save everything that I have into markdown files. So you can actually download like your notion documents everything into markdown files.

32:29everything about the business, everything related to the product, you can save that and what you can do reboot prompt, you can just select a folder or a bunch of files to say I want to a brainstorm strategy about business and I have a folder called business. So I copy that uh copy that to repo prompt, drop it in chat GPT, go for a walk and then open chat in voice mode and talk to it and that's that's just amazing. So like don't stay in sty don't stay freaking seated like use the tech >> more about voice mode. So the the one product of yours that you were building that I wanted to see >> that I wanted to use was um um voice to

33:10claude code or voice to cursor something like that and you're building that or why did you stop or like I just I just want to do what you said but instead of talking to my GPT and then feeding it back into my thingy like I want to just talk to it and you were building that and what the [ __ ] Uh yeah so I started building that. So it's called vibe flows. I started building this this thing and um I thought that my biggest problem is going to be initially just how uh latency latency was going to be my big problem and it's going to be like oh I need to generate speech really fast

33:43right but that stuff doesn't matter. It's the models themselves that are really slow and when I started building it I kind of realized that it's like hey my wind server is taking like 10 seconds to respond. Uh, I just I just I'm too ADHD and I lose my attention span. >> Realize that people in the same way that me and you you talked to me earlier like I hate Canva because I'm a shortcut guy, keyboard shortcut guy. And I don't prompt with fingers anymore. I only prompt with voice. But now I'm doing cloud code and all the [ __ ] that you're talking about here, but I have to switch back and forth from GPT's little

34:15microphone to translate my text and then copy this [ __ ] in. Can you build us? >> Yeah. So build the dictation part helps already. Uh so so I I want to answer your question actually. So during the process of building it, what happened is uh I went to websummit and then I lost my freaking voice and the last thing I wanted to do was speak to a damn computer. So I started typing again and I was like, "Wow, typing is kind of fun." Wow.

34:46So uh that's kind of that's a rock in the road there for me. Unfortunately, >> I want to get into some question and answers because this guy's like the source. So, thank you for revealing some of this [ __ ] earlier. Uh, does anyone have any burning questions cuz I got several. >> Okay. Okay, you're next. Um, this whole new thing about like um using agents plus GitHub to automate the whole software development process. So you vibe code up a piece of software, you commit it to git and then you start filing what bug reports or PRs and then you have your agent look for those things and then start to like prioritize tasks and you can essentially have the

35:23community filing GitHub feature requests or bug reports or pull requests and then essentially have my agents picking the good ones out of there visav my overall goals and automating the whole thing. Have you started to [ __ ] with this? Um, not quite because I feel like so running in the background is something I just started playing around with for simpler tasks. Uh, I feel like still running in the foreground. We haven't quite figured it out. So, I want to nail that that part of it first to be honest, otherwise you just be wasting compute uh in there and wasting time and not not from environmental perspective. I'm sorry.

36:01Um, we're like 1% of pollution in the world. I I'm an engineer and god damn it. Um so uh I mean just you're wasting your credits uh as and you're going to run out of them. So that's why I haven't used it as much. But what I do with Cyrus actually I can foreground the task. So it uses claude on my machine and I can just claim if it if I find it it was like stuck somewhere or like I just don't understand what it's doing. I can just stop it and just like cloud code resume the session and then start kind of querying and and vibe coding for a lack of a better term with it. So

36:39prompting it. >> I think it's we're going to you in one second, but I think it's so funny that I got you two guys back to back here tonight cuz you're [ __ ] like bleeding tokens all over the internet and he's like counting all the microwaves. I I'm like, "Hey, Michael, just uh make a leaderboard to figure out which one ran the longest." He's like, "That's so many microwaves.

37:00>> I want to be on top." >> It's actually what I love about this whole community is like we're excited about the opportunity and the possibility and we're also like, you know, trying to like hold it to account and figure out the best way forward or whatever. So anyway, homeboy in the back.

37:12>> I like I like the sliding scale and having both sides of it. Um my question um first time here is how breaking things it down into subtasks and adding those guard rails. uh are you kind of are you asking it to create its own guardrails because I've had situations where you ask it to change something ex you know seemingly small only touch this form and then all of a sudden it's ripped a whole section out. Um so are you are you building those guard rails for in in those tickets in those subtasks or are you trying to get the agent to you know work out its own guard rails?

37:56Yeah. So, sort of uh I just give it the subtask itself and sometimes I have these uh guardrail. I definitely had it um with Cyrus I give it a task to just implement a rewind button um on like a different screen, but it I was like, why didn't you go to the the player to like to look at it? It just made stuff up, right? Um so, it's still I think it's an unsolved problem, unfortunately. to be honest, I'm figuring it out. I think all of us are kind of figuring out. Uh I don't know about you, I get into fights with it sometimes. Um people said it's better to be nice, but I'm just mean to

38:34it. So, I don't know. Maybe that's one strategy there. I haven't tried that. I don't have the patience to be nice to especially to LLM. Um yeah, sorry I don't have a better answer. >> Hey man, thanks for coming and dropping your knowledge here tonight. And um who has a question? Niels, come up. Come up and ask it. Come hang out. Um, thank you, bro, for sharing tonight.

38:58Appreciate it a lot. Like, uh, >> thank you. >> Yeah. Uh, if you guys want to hire Michael to build your next AI company, he's not available. >> Um, for 100 million, I'll do anything for you. Uh, actually, I'll give you the Canadian discount. 50 million. >> That speaks to the moment that we're in.

39:17I'm trying to move to that moment myself where instead of building things for other people, I'm building them for for myself. Um Niels from Bridge to AI. >> Uh I just wanted to say I relate to the um the fact that you're waiting on the AI a lot. What I inspired by is the fact that you can have multiple things going at once cuz you know 4 years ago you could just like enter like a coding jam and you'd just be like fully like hammering like task after task after task after task. Now I find myself like sitting on YouTube like waiting for the thing to to load. My question is with the multi- aents

39:53um um how do you manage like having so many poll requests on your same app go at once? Because I find that sometimes it can change a lot of things all over the place and then merging another poll request it becomes difficult. So I I'm Yeah, that's the missing puzzle piece for me to be able to set off many agents.

40:17>> Yeah, I I I struggle with it myself. So it's not it's not just you like staying in the in the flow. It's it sometimes sometimes I'm having a great day and uh cloud code is is doing what I want. Sometimes they rate limit you and it gets dumb as hell and um you can't get anything done. So I I feel you. I think the I kind of run it on multiple workspaces in general like on my uh on my laptop I open multiple workspaces like each thing on its own port have like a command to >> it's it's beautiful. Come on.

40:54>> I skipped one of his slides earlier. Anthropic sent out an email to all users being like we're shutting you guys all down because of 5% of users. >> It's but will they be all working on the same app at the same time? >> Yeah, they are. Then you merge the pull request afterwards.

41:09>> I do merge the pull requests. So I tried to get them to write the smallest one. So um I I didn't show that slide, but I had uh the most I'll show you actually >> cuz uh that was I was I was so happy with this. Um so this this one here uh it fixes a bug with importing images into my uh into my app. It converts them to markdown in here. and it doesn't it didn't take it into account like having a title tag in there. It changes the markdown that get generated and generates it right next to the URL uh and it thinks it's part of the URL and it generates a broken image. So I didn't

41:48really look into it. I just told LLM, hey just like go do this fix this thing and it wrote like it moved this line from here to here. Uh and I realized I haven't looked at this code in like eight months. Amazing. And it fixed if you look at it, it added uh all images to it. I was just doing it on one type of figure tag in there. So that was that was amazing. And then it generated tests for it as well.

42:15So it generated four tests. I needed one test for it. So I just kind of like, hey, what the f are you doing? Um give me one test. And then I was like, okay, I'll give you one test. So I kind of have to babysit it a little bit still. Uh, and then sometimes I don't do a good job. Uh, and this thing happens. Um, and I have, it's so exactly like one of the speakers like took me one day to write this feature.

42:43Uh, and 7 days to like fix all the bug, the damn bugs that it generated like adding images to like EUB uh, and exporting them. >> So, uh, the critical thing is this number here. It needs to be like a hundred lines of code or something, 500 lines of code, like something tiny like a human gives you. Like if another developer gives gave me a PR like that, I'll I'll freaking fire them.

43:09Or um I'll I'll try to get them fired. To be honest, >> this is the longest I've ever stayed in a sauna. >> In a sauna. Oh, >> thank you. Oh, okay. Uh thank you for the presentation. It was really good. Everything's falling down. Um, I'm whole new to the AI uh community here. It's my first as my first um event with you guys. So, I really enjoyed it. I won't say that first. Um, as being like I I just graduated from computer engineering and still trying to find my way around. I'm still trying to get the hang of all AI and how to build everything. So what would be your um suggestions or like

43:56where to start the whole thing given the fact that you did like 52 times or more? >> Yeah, I I'll go back to that lever analogy in there. So uh you don't like to to move a pebble. >> He's finding his slides. I would say download Claude Code tonight before you go to bed.

44:12>> Yeah, do that too. But um you can you can also play around with lovable and like these easier tools. But um if you're if you're about to move a pebble, right, you don't actually need a lever, you don't need any tools. So if you can define the problems smaller, uh you don't need the tools. And then the more complicated your problems get, the more tools you'll kind of need and you you can slowly actually experiment with tools and add them to your workflow. You don't need to add them like wholesome to your workflow. And that's I think a lot of us have like FOMO. It's like, oh my god, I have to throw out my entire

44:49workflow and then replace it with cloud code because Entropic said that if I don't, I'll get fired, right? But uh you won't uh you should experiment with these things and see where it fits in your workflow and then grab the handle where you need it too. >> Controversial conclusion, man. I actually think you should throw away everything and download cloud codes tonight before you go to bed and give it a try. I think you said something earlier that really resonated with me is like a breakthrough I had this last month. myself is like um it's not just for writing code. The way you were talking about storing the whole business in markdown files and using the business

45:25plan, the marketing plan, all the research, that whole database I built tonight with 550 AI organizations in British Columbia I've built using cloud code and that is now the [ __ ] grounded truth database for me to make all sorts of decisions and like build all sorts of things on top of. So it's like I think you're wrong. Don't integrate it slowly one tool at a time.

45:44[ __ ] download every tool off his workflow list tonight and give it a prompt and give it a try. Give >> it a try. >> Yeah. Like I'm not saying don't give it a try. I'm saying you don't have to go all the way there. You can try it. Uh and if it doesn't work, you can kind of remove tools uh as well. And it's it's fine. There's nothing wrong with that.

46:04We're able to ship code really fast before like faster in 1995. So >> Michael Yagudaev, the coolest coder I know. Yeah. Thanks so much, guys.