s03 e08 Vibin’, Tokens, and the Future of AI Development

[MUSIC] Welcome back to the Big Cheese AI podcast. I'm your host today. [MUSIC] And I just want you all to know that Brandon just spent five minutes explaining how someone should build something unlovable and chat GBT. You don't even sell her own services and he's--

• How good does that feel, Brandon, when someone's like, "I wanna build this, can you build this thing?" And you're just like, "Do this?"
• They go do it then.
• Dude, it is so much better because everybody does. They're like, "Hey, I got this idea. I want you to build this idea." And I can now just be like, "No, you're gonna go build this idea. You're gonna go work through all the bullshit. And then at the end of it is gonna come out a product that's either gonna be great or it's gonna be dog shit. The responsibility now can become on the people who have the ideas of the product, which is actually a really exciting position because a lot of times they don't actually have to worry about that kind of stuff, but now they do. So you have to work through it. And so like this guy, he comes to me once a year with some big ass idea. And usually it's like me trying to tell him, no, I probably don't have time to build it. But now I'm just like, yeah, go build this dude. Like this is now your thing, go build it and prove it out. And then we can go make it amazing. So that's the kind of stuff that I'm just pumped. And lovable specifically, because he doesn't need to be a developer, he doesn't need to have an IDE running on his local machine, he doesn't need to know how to do MPM install or any of that, he can literally just go to lovable or bolt and basically just put in the prompt and get there. Now the reality though is that, and I don't know if you guys have done lovable stuff yet though, so I've got about five different lovable projects, and I can make it about a day a half before I ripped through my tokens for the month.
• Oh, well, we were just talking about that. So we were talking about that today. So I'm getting ready. Brandon and I are gonna be speaking next week at Elevate Ventures AI workshop. And Jacob, I was kind of bouncing some ideas off Jacob and you and some other people about the outline for the talk. And one of the things we were talking about was the honeymoon phase for pricing for these toolings. So for cursor, fastest company ever at $100 million ARR, they're burning cash. Opening eyes burning, billions and billions and billions in cash. But what's happening is that people are using their products and in the standard Y Combinator growth mindset, for startups, it's all about growth. It's all about revenue growth. It's not about profitability. And but at the end of the day, that trend has changed. And it changes quickly when you're burning billions, not millions of dollars. But so you can kind of take like, so Amazon did it, right? We all watched Amazon and with our clenched buttholes as Amazon like just continued to burn through money, right? That was their whole model that they were going to lose money at an insane clip to ultimately be able to eclipse and become like this huge profitable company. Clearly, you know, cursors kind of going down that same path because there is no doubt that when we were kind of talking about it on Slack where they're like, Well, maybe the honeymoon's over. I don't even think we're even remotely close to the honeymoon being over. 'Cause if you would go actually put in your own chat GPT, like API key, and you would just use like raw dog, your own API key, dude, you're gonna run out of tokens like in a day. So I wanna know how much you guys are ultimately paying for Kurshin. 'Cause right now I'm at about, I think I'm at like 80 a month is what I'm paying.
• Wait, 80 a month?
• Yeah.
• I thought, well, we pay 20 bucks a month per user. until you yell at it long enough that it's like, yo, you've ripped through all your fast responses and now you have to go into the slow lane.
• I haven't even looked at our bill.
• So yeah, I've bumped mine up to where I was like, all right, I'm willing to pay the 40 and then I got stuck and then I'm willing to pay the 80. That being said, with lovable, I was like, all right, I'll do the 20. And then I, that literally the next day I hit my, hey, you've run out of your monthly thing. I'm like, what the fuck is this? So then I upgrade to the 50 on Lovable, and I hit it there too. So Lovable is being much more conservative on their billing versus cursor. Cursor just feels, 'cause again, there's been times where nine hours, I will just be yelling at AI for nine hours, copying huge content. I mean, the chat context window on these things have to be just ginormous, right? And it's still running, and I'm not getting charged over the, like what it should be, it's gonna, they've gotta just be eating it. They've got to be.
• So the sweet spot right now is existing production code base, doing tickets, right? The stuff that you're throwing at cursor is not insane. Like you're trying to absolutely boil the ocean every single time you're doing, and you're building new. There's a big difference between using those tools to go build something that's new, because it's just like anything else. It's more expensive. It's got to do a lot more, you're constantly having to rip out and start over. But I think like we need to start defining the techniques in order to be more efficient because we're wasting a lot of the time with the cycles and we're being lazy, we're copying and pasting our message and just saying fix it.
• Right.
• You know what I mean? And I think that there's part of this is where you need to go in, you need to do a little diligence on your own, especially if you know what you're doing, which you do.
• Brandon.
• Dude, dude, there's one. Okay, this is just so bad. I spent, I had to be four or five hours. And I am now at a point where I'm just breaking and I'm just in my basement just screaming like a madman. Like we've gotten to the error loop, right? Where he's like, here's error one, go fix error one, we get error two, go fix error two, we now get error one. And I'm just like, okay, I'm fucking done. And I'm losing my mind. I eventually am like, I'm just gonna go look at what the hell the problem is. I go in, dude, it's a comment. It's a fucking comment. It's a comment, something weird and swift that was causing an issue, and I just literally just deleted it and it was gone. And it's that kind of stuff that people don't think about that when you start doing the vibe coding, you're gonna feel like you just won the lottery. And then you're gonna hit this one bug. And if you're not willing to yell at AI for the next five hours, or actually just get into the code to see what the hell is actually going on.
• Yeah, I like that option. (laughing)
• Well, but the problem is, is a lot of these people aren't gonna actually know the code.
• Well, that's why I don't believe that junior developers can vibe code.
• Yeah, yeah. Well, you and I had an interesting conversation this week about vibe coding, where you kinda got disheartened by some of it, right? Like again, I've now got, I don't know, I probably have like 10 projects, huge fucking projects, that all have gotten stuck at one point that I just cannot move past. And I'm too stupid to be able to do it just within the natural code itself. And that is a very disheartening thing where you spend two months building the saying you just can't get past this one thing, you kind of had a kind of a similar experience.
• Well, yeah, I essentially, I just hit that curve where we got the UI built out and got going and it looks great and I even demoed it to people and they're like, This is so cool. And then I'm like, you know, I kind of want this to work off. (laughing) I want to use, I don't, I don't, I want to use a super base. And then I want, I kind of want the sinking. And I want, 'cause people are gonna be using this in places where they don't necessarily have consistent internet. And then, you know, I go on this, I go on the endless churn. And you sit there and you go, man. And then I hit the absolute biggest pitfall I think you can have, which is you plateau up two plateaus. So you got your auth, you got your data, you got your UI or whatever. And then you accidentally haven't committed to Git and you just blow away your entire database schema. So literally all of the API, all of the UI controls, all of the functions that are interacting with the APIs that are built are now broken because the structure of the database is fundamentally changed. And I didn't write this code, AI did. So you think I'm going back and gonna go, and now, like you're mentally beaten at that point, 'cause you're like, I'm not gonna go. I built that part and it was working, right? In five minutes, you think I'm gonna go and manually fix this?
• I was just thinking about this the other day. So like, I had a very similar experience where I used AI, I was vibing, and I built something in like a day that it would have taken me two weeks. >> Yeah. >> Right? And just the sense of entitlement you get from that. >> It just creates that. >> I was the same way. I was like, well, I'm using, I was on Cloudflare, and I was using just regular SQL to talk to their D1 database. And I was like, well, I want to use Drizzle, because I want to make this a little bit more manageable. I want migrations, all that stuff. I don't really know Drizzle that much. I normally use other ORMs. But anyway, so I jammed all that stuff in there and all of a sudden stopped working with my code. And I spent like an hour trying to figure it out and you know, who knows what it was. It was some type error or whatever. But I just like after a while, I was like, I just wanted to quit. Even though it took me like a day to do something that would have taken me a week, I felt like hopeless.
• And it's the same thing that our investors and our managers and whoever is a, they're gonna be the same way. Like the, just be like the expect, oh, oh, that, that now only takes you a day. It used to take two, well, that's now my expectations.
• That's a good point. My expectations were completely not in line with reality. And I do think it's gonna take discipline and the same old stuff we've been doing for a long time is just gotta grind.
• Right. I mean, it's like the same thing as like, like when I had, you know, a 56K modem versus when I got cable, like, okay, why are my Napster files down? downloading, I'm on cable. Like my expectations change immediately. Like I don't go back and I go, "Well at least it's not the old days where I know you go, why is this not working?"
• So as a consultant, it's one of the things that I've struggled with as well, right? Because again, when you're doing things at an hourly rate, and now all of a sudden, the things that I'm able to do with a proof of concept where which would take me literally three weeks before, now I can do in a week. It's kind of like, how do you do this? Like, how do you position it? So at that part is really interesting. But one of the things that I wanted to bring up in terms of like talking about where AI just kind of goes off the rails was something that happened to me this week. So I actually started working on a native Mac OS and iOS application to manage your ghost blogs. If you guys don't know, we've talked about it.
• Oh, I've seen your posts on social media. You're following it.
• So, "Ghost" is an amazing open source. It originally came out as a WordPress competitor written in Node, and then they've now kind of switched to be more of a sub stack competitor, an open source sub stack, so you connect your stripe, and then you can have your different levels. You could do like an open source only fans if you want. Anyway, one of the things I realized is that they're editing. I wanted that the editing, the what you see is what you get is actually really, really complicated, right? And so I ended up going and saying, I'm going to use a React editor and I'm going to embed that into Swift. So I get my React thing all built and it's all working and I can import it as an iframe and I'm just using post message and on event listener. Only two ways that you're able to interface with an iframe. Okay? So I'm like, all right. And so then I tell the code, I'm like, "Hey guys, we've got the iframe here. Go ahead and import this into Swift to build a Swift editor that imports it in. You're going to be using post message and on event listeners for when things happen." And some of this stuff was working, but there's a couple things that weren't. And I'm just like, well, this is kind of strange. So we back and forth, back and forth. I'm yelling at him, like, why does it not work? It works in my test. It should work in this Swift. And then eventually I get into the code and realize that it completely bypassed the post message in the on event listener. And it was literally writing JavaScript dynamically and injecting it into the DOM to try to call local variables and local functions within the editor. So it, because I got so mad at it, it said, okay, I've got to just try to figure out how do I solve this problem? So he just stopped saying I'm a fucking asshole and it actually works and it did, but it did it completely in the wrong manner. And these are the types of things that people can't sleep on. Like, I mean, this is such a horrifying example of like, you have to validate the code that's coming out of these AI systems. Yeah, I mean, the fact that not only even-- you're just talking about the app code. There's platforms out there. And ones that are going to be choices that AI uses to build their apps, like full stack applications where the infrastructure is in the code, the database schema, the migrations like you talked about are in the code. The references to secrets and API keys are either at least referenced or written into the code.
• If you're not using cursor, ignore right now, understand cursor knows everything that's in your .env files. But people don't understand that. Like, there's a cursor when it indexes, it indexes everything. So your .env that has your open AI API key in there, that has your secret key for Superbase, that has all of your different secret keys, that's getting indexed and processed through cursor. And that's why it can do those things. So you need to have the developer community just went and rolled their keys. After they just pressed pause on this podcast.
• And you should. You absolutely should. So cursor ignore is just like a get ignore where you can go in, you can put in. So put in your .env, put in any of your secrets, leave your .env-example. So cursor knows, oh, hey, I see that you've got an actual example of an ENV. So I'm assuming that you have a real one.
• Oh yeah, I mean, the amount of secret keys and private information that cursor has in their system, I guarantee is all the characters.
• Well, you know, when you ask OpenAI to write, it'll just like paste a key into your code. You're like, "I wonder where you got back."
• Who's key is that?
• It'll like work.
• Dude, there was a guy that, and I posted this one on LinkedIn where a guy, and I don't think we've talked about it, but he was a vibe coder. He was a designer and he was doing some vibe coding stuff and he was building in public. So he's on X and he's like, so I go and I push my application and now all of a sudden everybody's using my API keys. Like I'm just getting hacked left and right. Like so please just everybody has to be honest about what this is. It's again, AI is gonna be able to build you what you want but you have to remember that AI wants to make you happy. It does not care about making you secure. does not care about making your app secure, it wants to make you as the owner happy.
• But isn't that the opportunity for cursor? So one of the reasons why chat GPT was so successful is what you've always talked about, the pre-promps and the processing before, and not just going LLM. I think cursor is probably gonna, and they update all the time. And they're the forefront of this, so I know we're just literally a cursor fanboy, but the point is, I guarantee that there's going to be a migration more towards the lovable, irreplet models where they're like, okay, what stack is this? Okay, before we do this, like what color, like maybe I should force some clarifying questions or maybe I should, you know what I mean? Like, 'cause I think the conversation right now is just so open. Like you can literally just be like, you are my whipping boy. If you're interested in, and I think that's part of the reason why Brandon loves this so much is because he wants somebody to say yes. (laughing)
• Just listen to me, please. (laughing)
• It's the point there is like, is I think that there's gonna be a, they don't wanna burn all those tokens. They want client customer happiness.
• Right.
• They don't want, they probably don't want their average customer, and maybe they do, to spend more than they can afford or whatever. I may be, but I think that goes, and going back to that, I think the more valuable they make their product, and these people make their products, especially if it's a true companion to a workforce element versus just being a one and done type of thing, the true companion AI tools, they're gonna start justifying very very, very, very, very large hefty pricing models.
• So when they do that though, and I don't disagree because the moment that open AI goes out or that cursor goes out, right? We all get that feeling of like, I've gotta do this by hand.
• It's unbelievable.
• I don't wanna do this by hand.
• It's like going from, it's like going from, you know, JavaScript to Cobalt to Assembly or something like that, right?
• Right, so but, but here's the argument for it, is what they're gonna do, is they're gonna realize, oh hey, we've been supplementing the cost of this, and now all of a sudden you guys are gonna be paying three, $400 a month. All of us will probably be like, fuck, okay, I'm willing to pay it. However, then there's gonna be those bastards who are gonna be like me, who are like, I've got a 96 gigabyte Mac studio at my house right now that I could potentially run an optimized deep-seek model that's specifically designed for code. And at what point can we just start using that?
• That's where cursor needs to create their, they don't have a moat. So cursor-- - They have two models. So they have cursor.
• So there you go. - There's a small--
• But that's the point. - Yeah. their model, cursor will have a mode. Once they make interacting and building, interacting with their product and building applications, more likely to be successful than just using the latest and greatest coding model. Because as we know, we just talked about it. Don't mess with my secrets. Don't rewrite the database. Where are we at in the project plan? No one talks about this. What's the, how do I, how cursor needs to know everything that you know? Yeah. (laughs) They need to know where you're at. They need to know your thought process, your strategy and where you're at in the project. You know, they need to know so much, and I think they will. Right, yeah. So the MCPs, which is, I think we'll talk about that next week, and I'll probably end a little bit more. We're talking a little bit today, too. We're talking about today, but this is though. So I finally have started kind of incorporating that into my development workflow with Incursor, now that they've got to be able to-- How have you implemented MCPs in the development workflow? So they've got now where you can just go and add MCP servers just like you can with documents. And so the two that I've added is the Brave Search and then the browser extension. So there's the, I think it's called like browser MCP.
• That uses the puppet, is that the puppet year one that Chad was showing me? It can launch browser and do tests.
• Exactly, sorry. You can basically, what's nice about it though is what I'm finding is that YOLO mode that you can get into where you can basically be like, "Hey, listen, here's the problem. "You keep fucking this up. "Like, we've been battling this for five years, "or for five hours. "I want you now to just go write something "that you can run in the browser to test "if it actually works or not, "and you're gonna just keep iterating on this "until it eventually works." And so then I just set that thing up, and I just sit back, and I just let that thing run. And now it's going, and it's gonna keep kind of testing itself until it eventually comes to the solution, and then I'm like, "Okay, now we're here. Now again, you're gonna have to every 25 token or every 25 tool runs, it pauses in cursor and you have to hit the resume. But other than that, you can pretty much just sit there and not fucking wrong.
• I don't think there's a single person in America that burns more tokens than a smotherfucker.
• Dude, no, literally part of my prompts is like, motherfucker, if you need to go burn a billion tokens to solve this fucking problem, I do not care to go burn as many fucking, like I lose my mind. But yeah, I guarantee you cursor hates me. (laughing)
• The Amazon rainforest hate you.
• Oh, I know.
• Well, so I do kind of feel bad about it too, right? Because the reality is, is that I do realize that the inefficiencies of LLMs, like, I mean, the amount of energy that I have burned, just to solve these stupid problems, it does kind of hurt my feelings. And that's where like, I like the people like Data 317 and some of these other places that eventually, I think that it's like, God, I don't know, it just seems like where it's just so horrible.
• Oh, well, you're constantly trying to see what this stuff can do. You're like stress testing AI 24/7, right? That's pretty much what you're doing. But there's more practical use cases here. (laughing) I mean--
• How dare you, sir.
• And I think I go back to, there's an efficient way to leverage this stuff and have the right mindset about it in terms of getting more out of each request and understanding the limitations. But there's also nothing wrong with YOLO. You're doing it for a reason because eventually you know in your mind that this is going, it's gonna just click and you're gonna figure out, you're gonna be the one that figured out how to get it to do it. You're just going about it with absolute un-inhibited force.
• No, no, not uninhibited. A lot of alcohol and rage.
• But I think like there's a lot of value to just throwing everything against the wall with it and seeing what sticks, 'cause that's how I've learned to like, all right, what is a good prompter and snippet of code that I can, what's reasonable, right? But I will share an embarrassing moment today where I was trying to center a div and make it full width image and I already had the CSS that would have made it work. So I like highlighted it and I was like, "Hey, do this." And it just said, "Oh, one second, churned for like 15 seconds and printed out the exact same code." And then I realized I had an issue with like the container, like a container up. I had to highlight it that. So I'm like, "Oh, that just took like a minute of my life to be, you know, while it's churning, you're like, I'm so efficient." (laughing)
• And then it ends, and then it doesn't work. And you're just like, "Oh, geez." So, dude, I lose my shit. So, like, eventually it's like, "Oh, I see the problem." I'm like, and I stop it. I'm like, you do not fucking see the problem. You've been telling me you see the problem for the last 20 fucking minutes. You know what's the problem? I guarantee I bet you a hundred dollars that you do not fix it on this time. It's like, all right, let's fucking go. And then it tries to do it, and then sure enough it won't.
• Dude, yesterday I don't normally get a snarky or mean as you do, but yesterday it solved something. And I was like, how the, hmm, you said that you solved it, but I'm looking at the result, and it's not fixed. Can you explain yourself? And it was like, you know, the typical, you're absolutely right. Wow, you're so smart. Thank you for pointing out my inefficiencies. I'm a dumbass.
• It's so frustrating.
• Oh. (laughing)
• Yeah, so that drives me crazy. It absolutely drives me nuts. Are we good?
• We're good.
• Yeah.
• Just make some tweaks.
• Just mid show, just adjusting the settings here. So I can't hope but notice the laptop that's sitting next to me here is Sean. That looks like it might have been delivered by Ace Ventura. (laughing)
• Dude, the corners are like smelted. Like you're turning them from aluminum to oxygen.
• I treat my MacBook like you treat your AI agents. (laughing)
• Crazy, crazy. Yes, so anyway, that's where we've been. But are you guys using any of the other models now? So, cursor swapped to their new, like where they do the automatic stuff?
• Yeah, I just automatic.
• So I do automatic usually until it eventually gets me so pissed off that I'll go to think.
• I think that there's a problem with that that I don't like, which is you'll be mid conversation. You can tell just switch to thinking and I'm like, why?
• Right.
• And like you just literally churned and it was way faster but like it went through like six different things really quickly. I'm like, I didn't want you to do that. I still think for me, for production code bases that base, sonnet latest, you know, model is really good.
• Have you guys used the Gemini 2.5?
• I'll kick it around every once in a while, but like, I haven't had that much success with it.
• Oh, I love it.
• Oh, do you?
• Are you using it like?
• Well, I use it--
• This brought you by Google.
• I mean, I use it in Google. I don't use it in cursor.
• But you don't use it in cursor.
• Only because the last two times it said it wasn't available for some reason. I think they had some issues with their API, but--
• Well, Google is releasing a lot of stuff and they released the agent date. They're kind of almost like an MCP type thing. What's it even called? It's the dumbest name ever again.
• It's A to A. So agent to agent and it's a new protocol. So very similar to what like MCP was where MCP model context protocol is a new standard protocol to allow people to basically build tooling that you can incorporate into your AI. So the great thing was, and I was not convinced of this, but OpenAI has embraced the MCP. So I'm very hopeful that this AI community might not just be a lot of the twats that the previous technology empowerment people were like, we're not gonna share our social media stuff. We're not gonna have an open standard, right? That OpenAI, I think, is like--
• Well, they have no reason not to, because the MCP piece is really pretty simple. Like the way that--
• Oh yes, but it's not something that they need right.
• The open, like if we think about like what open social was trying to do, right? Where open social said, hey, let's have a standard so we could have interoperability, we could have these things, so we could have this kind of like, this true kind of panacea of social media experiences, right? And OpenAI was like, nope. And then you got Twitter, there's like, nope, we're not gonna participate. So I was, as the old man that I am, I was prepared for OpenAI to not participate.
• Well, that's true. Open AI's, what's their other option would have been to create some proprietary protocol that every single product now has to leverage and go down that path for two seconds. They would have gotten passed by. But MCP's great because it doesn't violate anything that for Chatshibiti that they do. It's the best use case is I'm sitting in a chat with an, could be Chatshibiti, could be Claude, could be whatever. I'm sitting there chatting and instead of just chatting with the LLM, I'm chatting with the LLM, but I've accessed other stuff and to do other stuff. Yeah. So one of the things that I've been thinking a lot about and of where this goes is that companies, eventually these large enterprises are going to have their own MCPs. So one of my big insurance companies, I'm kind of thinking, what is that? Well, they're going to have to have their own MCPs because guess what? You can't just go to Farm Bureau and plug in five fucking GitHub repositories that are plugged into all your data and just trust that that shit's going to be secure. Exactly, exactly. So, but now imagine being like, hey, and the setup of the MCPs is still too new right now. It eventually becomes a point where you like use OAuth or whatever and be like, hey, add this MCP server to my Google or my Gemini or my chat GBT. And then you authenticate against your Farm Bureau, account. And now at any point I can be like, hey, I just had an accident. I want to file a claim. Right. And now the MCP goes, Oh, hey, there's a file claim tool. Great. I'm going to go kick in the future of basically everything. I think I really think it is. There's there's the future of the next 10 years of software development for enterprises is going to be to stop building UIs and forums. Yeah. And just buy. Yeah. So generative UI is the other piece to it. Please take a picture. Please upload a photo of your thing. Oh great, I got that. It's in the system. It's in the claim. But that upload button, right? So that upload button. So right now what we have is we have this chat kind of, and we're starting to see it. But it's in the chat. It's in the chat. But we're starting to see where certain UI elements are bubbling up in the chat interface right? Sure. More and more complex. We're going to get to a point where you're going to have the Ux paradigm is the chat itself. It is. But I do think that we transcend the chat to where you actually have full UIs that are going to be very much customized towards the end user. So when I launched my Indiana Farm Bureau mobile app in five years from now, it's going to be a UI that's only going to be for me. And it's going to be specific to the policies that I have, to the things that are going on in my life, all of it's going to be customized. So this idea of even building UIs for personas is going to be-- no, you're going to build a UI per user. It's kind of like where my head's at with it. Yeah. I agree. And I also think this agent to agent and MCP model is the next-- it's the best thing since sliced bread. Exactly. I was building in Cloudflare last week, which I don't have a ton of experience with, but I've always followed them. And I got the Cloudflare MCP. And it's pretty cool to just be like, hey, I want you to spin up. but I think it was, what did I do? Oh, I did a key value store, 'cause I just wanted to build a quick caching system for the search thing I was doing. So can you create a key value store? And I could have gone to their docs and they have a very easy CLI with Wrangler to do that, but it just said, yep, let me call the tool real quick.
• Where are you doing that? Enquod? - Enquoser.
• Enquoser. - Oh, you were doing it? So you went, you set up the, encursor, you added the CloudFlare MCP, and then you were just doing it through there. I assume we could do the same thing through Claude desktop if we added the MCP MCP there as well.
• Yeah, I have it on Claude desktop. I have it on Claude desktop. That's where I started, because it was a little bit easier until a recent update with cursor to install MCPs on there. But I installed Jira as well. It was really, I had a good automation going with like puppeteer and Jira where I was like, I'm gonna explain to you what the ticket is and where it's happening at. I want you to go take a screenshot, capture what's going on and create a Jira ticket in a tax that screenshot. It got 90% there, but the screenshot was locally stored. So I had to actually download it and upload it to Jira. The next step would have been to like move that into a product and actually like upload the image somewhere publicly so that it could have accessed it or whatever. But I was like, damn, that's like, (laughs) the ticket was fantastic too. I was like, man, this is, that's where I think it's really shined for me. Like we talked about all our use cases and I keep going back to writing better descriptions, specifications. - Oh my God.
• I did an incident report last week where I just spit out everything that happened. I just recorded, I used the whisper flow, recorded exactly what happened. I said, "A client found this, Ken did this, I did this." But we went through the whole process, talked it out for two minutes with Ken, one of our developers, put that into chat, or actually I used Jim and I, but it doesn't matter. And I said, "Give me the instant report." And it was fantastic. Like the client was so happy with like how thorough we were. And I'm like, yeah, this is the stuff that I don't, like nobody wants to do that. I don't want to, take me an hour to build that report. And I had all the information, I just had to give it to someone to help me format.
• Exactly.
• You know?
• Yeah, that's a whisper flow that you recommended.
• Apparently a whisper flow.
• There's a little shady.
• No!
• You sold that too?
• Yeah.
• I did, I was on Reddit, and it was like, according to the Reddit people, there's some privacy issues with the whisper flow, but dude, that application, it just works, man. You hit your, the only thing that I hate is that it uses the emoji key, which I actually use a lot. But you can choose to see. You can assume that, but you just hit the button and you just start recording. And the translation that it does is wildly accurate.
• It's great, it's great.
• The commas, the pauses, the spaces, everything. I haven't did a job that right.
• That's the value of it.
• My favorite thing, and Apple will get there eventually, hopefully one day, but I was doing my text, I'm like becoming an old person. I was like doing, I was speaking to text in my car, right? 'Cause I was like, oh, my phone's up here, I can do this while I'm driving. And you don't always have like full thoughts, right? So you pause and you re-say the same thing. And like, I think Whisperflow does a really good job of, and it doesn't--
• I wonder how it does it.
• Yeah, I don't know.
• Well, it's not their product.
• Because it's not raw dogging. It's not raw dogging the Whisper model because the Whisper model wouldn't do it. So it seems like there's a little extra spice that they're adding into this to be able to get that level of--
• It's very good.
• It's a very good--
• It's a China, it's so true. (laughing)
• Well, if you send it to China, gonna come back with 125% tariff. (laughing)
• Shit, fuck.
• Mike down, Mike down.
• Sorry, Mike down.
• All right.
• I get a picture of you falling.
• We're gonna mute that. All right. - This is great.
• Don't touch it, don't touch it. Don't touch it.
• So I had a client that's using Claude desktop and setting up MCPs and he just was like, "I cannot deal with this." Like, it's cool and everything, but like, this is just, and he just pointed to the repo that it was that the puppeteer repo that he was using to, 'cause he was just basically, he set up a thing where he just goes to Cloud desktop and he tells it to run any test on their app. And it launches the browser, it logs in, it does tasks for him. And he was like, yeah, just does it. He was like, but this is absolutely horseshit security wise. He said, look at this thing, it's just pointed to some random GitHub repo. >> Oh, dude, no. The security issues that we're going to run into now is just going to be absolutely horrifying. So we've talked about the one vibe coding project that I was on, that it like basically leaked. So the data, again, AI does not care about security. AI wants to make you happy. I'm just going to keep repeating this over and over and over again. Because when I get angry at AI, it will solve the problem, but it will solve the problem in the most insecure way possible. And so again, we were using Lovable, and it was like, hey, go show me all of my music that I've uploaded to be able to select from this list. And it showed me every single user's song that they've uploaded, because it did not apply any kind of SQL logic to basically filter by the user ID. These are the types of things that are going to be lighting the world on fire for 2025 and 2026. as some young little hopper comes out and like comes up with this new hot item that everybody's gonna use and All of our data is gonna leak and it's going to be a fucking nightmare. You know what else is pissing me off Is I've seen these these 23 year old kids that are like better They're making these videos and they look so confident and they're like this is our new AI app and then they show a video of it It's like it's like literally hello world but it does something automatically. And I'm like, they're like, and then they have some really big click bait title to the video, they're like, automates your entire accounting system, day one, no problem. And literally the video was running a headless browser, typing some shit into a, taking a PDF, that downloading, going to a Gmail account, downloading a PDF, then uploading that PDF to a Google Drive, and then type being data into a Google sheet. I'm like, nobody's fucking accounting system in the entire world is manually downloading PDFs from Gmail and uploading them to a Google Drive and manually entering the invoice details into Google. You just created nothing. That's not a product.
• It's gonna be so bad. It's gonna be hilariously bad that the stories that are gonna come out of the absolute failures of these software applications is gonna just be glorious. So do not be misled here, right? Like as you guys are building these things and you're trying to build these things, understand the risk that you put yourselves on very well can get you and your company sued, can get you and your company to go bankrupt. Like these are not things that you just kind of like, "Eh, we'll see what happens." Like you're just letting these, like there's real ramifications that can come from releasing really shitty software. Release shitty software, I don't care. Like I think it's gonna be hilarious. I'm all for it, I'm gonna see it happen because we're gonna learn a valuable lesson from all of this. But for those who don't wanna be embarrassed, you've really gotta validate every single thing that's coming out of these AI systems. Like you just have to and it's gonna suck. Nobody wants to go review the code that comes out of an AI system.
• Well yeah, that's the kind of elephant room. (laughs) Also like back to our original topic of if AI can make me so if I can do something that's gonna take me a day in half an hour But that's without checking my work. What if I check it and it takes me four hours. Yeah, well still half the time But nobody Not when you've got the expectation of it only took me half an hour to do it. Yeah, yeah That's why I think that there's a there's a protocol here There's a process that you I've seen work in production code bases that is incremental Where the changes are incremental the tickets are incremental and you plateau up you layer up instead of trying to boil the ocean every single time You said it best improve 1% every day You know and AI can make you do that like AI can can help you kind of continually improve But but it might be to pull it back a little bit right like like have some restraint I know you I know you think you can go boil the ocean now that you've got AI and that it can do all this Maybe It's the same shit though. It's like if I went and went to Claude and said what's the context window now 120 I said go write me a book or write me a 30 page paper on you know China US relations It might be okay, it might not be the bad bad, but it also might be dog shit too, right? and unless you read it.
• Yeah, but what if I gave it 50 points that I wanted to make and the references to it, that's gonna take me a little bit of time. That's the same thing as writing specs or being more meticulous and going feature by feature within your application and really thinking it out and using AI to supplement, which I do think it's really, really good at, is augmenting and filling in gaps of already pretty good thoughts to begin with. But if you're just like, go write the 50 page paper, It's gonna be 50 pages of shit, you know?
• Dude, so-- - Or mediocre stuff.
• As we're kind of working through this rally next week, the thing. So my thing, I think it's gonna be the 10 commandments of vibe coding. We're basically like, here are the different things that you just have to accept as kind of that reality of what you're actually getting into, and that's just a huge piece of it. Where again, you have to, you have to, that thou shall start with like an actual like projects back.
• You need to write this right now.
• Right.
• And then put it on the internet.
• But use chat GPT. So here's my thing is use chat GPT and D-C.
• That's what I do.
• So, well not fucking using deep C.
• No, no, no, no, no, I'm sorry, deep research.
• Oh yeah.
• Deep research, deep, deep something. We're going deep. That basically we use the deep research to say, hey, let's generate my first product spec and then it's all the documentation. The best results I've got, vibe coding, is when I talk to ChatsheetPT and tell it, "Hey, our output here is going to be a prompt for the first input to cursor to build a project." If you do that and you have a good conversation there, you're saving yourself hours upon numbers. That's the levers. You have to get somebody's going to literally write the book on this. You see it all the time. You see those clickbait head-ons. AI is changing every day.
• Here are three tips to get the best out of chat GPT. And they're all bullshit.
• Yeah.
• But there is somebody that, you know, there, I would, I read them. But, you know, there's something to those things where it's like, there's a reason why you put a little bit of work and nobody wants to do that. It's the whole point. You wanted to do it in five minutes. You had this expectation of like, I want instant gratification, I want to open it and we don't see it.
• Yep.
• But it's like, take the foot off the gas for a second because I think that you're gonna save in the long term, just like in the real world, with real people. (laughing)
• Yeah, well, it's like any software or tool is there's a learning curve too, right? If anybody does woodworking or anything, like there's all sorts of tools that can make your life a lot easier, but jigs and whatever, they take time to set up and to use properly and you have to learn them. It might take you four hours to set up initially that you could have done the task in an hour, right? But now every time you use that thing, you're gonna save yourself 45 minutes, right? Or whatever. There was something, oh shit. I was gonna say something about what you were talking about earlier about moving on. (laughing)
• Well, hey, so we got big, next week, we are going to, at this point--
• Are we recording any of this? Are we gonna record any of the, can we do that? Can we have like somebody, can we have Job show up and like actually record what you and I are doing. 'Cause like, I think you're gonna present and then you're gonna go on the end.
• We need to post that on YouTube.
• Yeah, and we need to have it be under the big cheese brand.
• Yeah, hell yeah.
• Okay, so we need to have that happen.
• Right.
• So how come, well, all right. We got some work to do. But next week, hopefully we'll be posting, we're talking to 160 people about AI. You know, we're gonna be talking, we got a couple big time, awesome data scientists talking before us. So hopefully we don't, We sound like we know what we're talking about by the time we come up there. They're just going to shit on. I will be talking about managing developers in the age of AI and then Brandon's going to hit. I'm going to introduce Brandon. Yeah. Um, I'm going to tell people to check out our podcast. I'm going to tell people to check out the vibe coding competition we're hosting at rally in September. That, that, that website just went live. Yep. Do we probably don't have the off the top of our heads, the rally. Oh, just people just, just Google rally vibe coding competition. It's the number one reserve. Right. And so right now though, I mean, the competition is 10 grand. So like if you guys are actually listening to this and you are an interesting participant, I mean, there's a $10,000 award that can come. And I think there's 12 or 16 teams. I forget. I think it's 12 teams and I'm going to be reviewing all the applications. Awesome. And figuring out what are the teams. Brandon, can you repeat that one more time? I didn't have the camera on you. Oh, OK. Basically that we've got the vibe coding competition coming up on the. September 25th 24th at the rally conference search rally vibe coding competition $10,000 winning prize 12 teams are going to be selected you have to apply for your team You don't actually get to know what the the competition is going to be about until until later on so come prepared and Prepared to kick ass so we'll obviously be talking about that next week Yeah, so looking forward to it. It'll be an interesting week I think a lot of a lot to talk about when we come back next Because we're gonna have a lot of A lot of people are either gonna people either going to I think there's a lot of different emotions that this all draws from people And I think that there's a lot for us to learn from other people - yeah, so so we'll see This has been the big cheese AI podcast with your hosts Brandon Corbin Jacob hello all right you guys have a great week We'll see you next time on a big cheese A.I. podcast. (upbeat rock music)