It’s Always Been Vibes

Hey, everybody. Welcome to New Jersey. This is where I'm from. It's a very good place. The Statue of Liberty is actually in New Jersey. Did you know that? Yeah. Yeah. Don't forget it.

So I'm here today to talk about vibe coding. For the past couple years, you've seen a lot of this stuff. Can I see a show of hands who's tired of this shit? I actually can't see any hands because I'm blinded. Yeah, it's super grating to see. The AI landscape is hard because there's so much noise going on, a lot of grifter-y types.

It's really hard to extract what is actually valuable versus the insane amount of noise, which I suspect is mostly bots. But also some troubled young men who are very vocal on the internet about how much they love AI. It really boils down to two camps. So you have super excited about AI virgins. Telling people they're getting left behind. Next week is, this week's going to change everything. Gemini 3, everything's going to get changed. 300 MCP servers. It's cranking out landing pages. And then you have the uses AI but doesn't talk about it Chad. Just hits tab in peace. Doesn't tell anybody about it. Chips. I guess it's cool. But that's really what you have out there. And they've been called vibe coders almost as a slur, which they are.

And they've been called vibe coders almost as a slur, which they are. But I'm here to talk about how vibe coding isn't all that bad. And it's actually not that different from how we code.

So vibe coding, at least the term, got brought into existence from this Carpathy tweet. This is a real TLDR. So he's basically introducing, saying, you know, I'm just going to give voice instructions to the LLM and have it build stuff. And I'm not going to look at the code. And he's just going to be like, yeah, let's modify the padding. That's great. And really just vibe with it.

I don't think what's taking place here is actually that different from what we actually do for a living in general. Think about what, like as a job, what we do every single day. Like the actual individual tasks. It boils down into a couple different things. You're Googling shit. You're stack overflowing shit. You're reading docs. You're typing code. You're thinking. And meeting with people and iterating. And what these LLMs are doing is taking care of a lot of that research and typing for you.

But you're still thinking and you're still meeting with people and iterating. You're still architecting. One of the things that I like to compare it to is, and it was less friction-y at the time. Do you remember when Prettier came out? I remember people were a little pissed off about Prettier. Not that pissed off. But I remember the first time I tried it, I was like, this is the best thing ever. I'm never going to format my stuff ever again. This is incredible. And I guess if LLMs did a touch better of a job, it would probably be like that.

It's a little annoying sometimes when they mess up. But I think that it's not going to do all of our job for us. And the emphasis on the thinking, I think the real value here is what you know about your domain. Your domain knowledge.

I had this tweet that semi-resonated that said the science to prompting is knowing what the fuck you're talking about in the first place. Right? This really proved true the other day when I was talking to my buddy, right? And he largely vibe coded like an analog synth thing on the web. It was really, really well done. I was like, damn, this is sick. But as someone who messes with analog synths, as well as the web, there was no way he could have got this over the line if he didn't have any experience with analog synths or the web. And I think that domain knowledge is why he was able to put that together.

But you can't just LLM how to perfectly design a modular synth. You kind of have to know. I saw this tweet the other day. I don't know if this is bullshit or not. It said that in the investor updates, they're looking for people who actually know how to program, like fundamentally. Because it's so valuable when you're trying to tell this thing what to do. Here's some rules that I think help. First and foremost, know what you're talking about. That really helps get the deterministic result. And the second most important one is read every single line that it generates. You have to read it. Even if you're vibe coding, it's helpful to read it.

And the second most important one is read every single line that it generates. You have to read it. Even if you're vibe coding, it's helpful to read it. It works, right? You're responsible for it. So that's motivation enough to read it. But as a side element there, you don't want to be a stranger in your own code base.

Giving high-quality context and examples is another one. I think you'll talk to a lot of these guys who are really wrapped up AI guys and they have newsletters and they're selling books and shit. And they're going to tell you you have to do all this, this, and that. I used to have a friend who coined a term for this. It was the funniest thing I've ever heard. He called it LLM astrology. It's this bullshit that doesn't really help.

What really does help is really relevant TERFs, contexts, and examples. Another thing that helps is not trying to one-shot it. Don't draw the whole horse. If you're going to use this thing to code, use it how you code. Say I'm going to build this site. What am I going to do first? Build the whole site? No, we'll make the header. So you tell it, make the header. It's the easier thing. So iterating through piece by piece and kind of staying in control.

One super important thing that I don't even have to tell you is to question everything it tells you, because it's a filthy, filthy liar. The funniest thing about the vibe codery guys is that I don't think they know that. You see them, they'll post stuff to Prod. They'll be like, look at this sick SaaS platform. It's going to make a million dollars. It's like your entire API's mocked out. Did you even check that? It's like, it's lying to you. That kind of feeds into the verify it actually works part.

So really at the end of the day, if you're a working engineer trying to actually use AI in your workflows, whether vibe coding or seriously, the most important thing that you want is determinism. Because a regular LLM, it's a non-deterministic system. Yeah, OK. I'm too wild. I knocked it loose. The way that you get that is A, through a high quality context, which is coming from you and your prompting via domain knowledge. If I'm telling it to write something in Canvas, I know Canvas. I know the nuances are on Canvas. I could tell it, do it, but don't do this. And then also examples and documentation.

The other part is high quality tools. I think a big mistake people tend to make is thinking that the LLM will just go do stuff, and it won't. One of the best ways to use it is to have it kind of act like an orchestrator where it can decide what tools to use, and these tools are literally just functions. This is code. It's using code that you're writing. So you can do things like if you're relying on the LLM, you'd say, you know, how many RRs are in the word strawberry? And that's like the hilarious example, and it doesn't know or whatever. Because it's just doing next token prediction. Well, not really anymore. Most of the things are agents now. But in an agent setup, you would say, hey, how many RRs are there in a strawberry? And you would provide a tool like strawberry word count tool or character count tool. And it will call that, and it gives it the opportunity to go call some function to get the answer.

So while vibe coding has been kind of like a pejorative term, I'm actually going to make the case for why it's awesome and why I do it not all the time, but sometimes. I'm going to show you some of the stuff that I've vibe coded. So you think about your day. Right? You have a lot of things on the schedule throughout the day that you're doing. And if you want to code, right, you're going to have to take a fat block of that day, depending on what you want to build. Historically. Right? You have to sit down, lock in, do some coding, all right? Get up from the computer. You're done.

You're done. And that's just a block you're never going to get back for better or worse. But now, you can kind of code in between with these vibe coding agents. As a matter of fact, you can code adjacently. And if it's something that you don't really care about, does it really matter? One of the examples that I like is this thing that I did, and it all falls under personal software. This is shit that I don't really care about. But stuff that helped me.

And a lot of the time, when I code to solve a problem, it's when I'm angry. So I am at the happiest place on earth. Or is it the most magical place on earth? I don't remember. I'm cooking eggs in the kitchen and my wife's like, where are we going to go? And what's the schedule? She's yelling at me or whatever. And I'm trying to use the Disney app. And I'm sorry if any of you work at Disney, but the app was slow as shit. I'm like, oh, God. Loading spinners. And it was driving me nuts.

So I only had my iPad with me and a phone. So I set up a Linode. And SSHed into it. And set up Cloud Code. And I built my own Disney app. Just for me. And the funny thing about personal software is you can do a bunch of goofy shit that you would never give to other people just for you. So here, this was actually pretty sick features because, A, it was super fast. And I did this while cooking. I just go back over to the iPad. I'm like, I'll make this like this. And I just go back to making scrambled eggs. It's very partial throughout the day.

It's very partial throughout the day. This thing has the times, the walking distance. It was amazing. And at the time, I was doing carnivore diet. So I was relying on turkey legs for sustenance. So I actually built in this feature where it would show me the nearest turkey leg location. And that's a ridiculous feature to put in any software that you intended to release to the general public. But when it's just for you and it's just my thing, you can have it. And it's actually hilarious.

This is still live, and it's at mousketool.ai. I can't believe I have that domain. When if you're going to Disney, and you like turkey legs, shit's sweet. But that's like half of vibe coding, right, because I set up a server, and I was cloud coding. What about real deal vibe coding? I'll tell you how to do it, and that's with a Repl.it. Repl.it is awesome. And companies like Repl.it and Bolt are well positioned to be vibe coding machines because they were just code platforms prior. You would just write code there the old fashioned way, uphill both ways, and deploy it.

I used to go on there and write QBasic. But now it's set them up. They have the architecture and everything in place where they add an agent. And now this thing, I couldn't find the tweet, but it was awesome because I had an Oculus or a MetaQuest 3 on. You can control with your hands now. You don't even need the controllers. But you have a browser. And I had Repl.it up in the browser, and I was on the toilet with the VR headset on, vibe coding while on the toilet. It was awesome. That's how low touch and vibey this can be. So I was playing video games, right? I hunt here in New Jersey. There's actually woods, but they're south. And the other time, when I'm not mad and I want to build things, I say, oh, that would be fucking sick.

So I'm sitting here, and I'm playing this. And I'm like, oh, you know what would be fucking sick? If I had this map with markers, right? And it can help me with the hunting. I say, yeah, I should do something like that.

And then, believe it or not, I was playing another video game. I don't know if you guys have ever seen this. It's called Ready or Not. It's the sickest game. You're a SWAT team. The loading screen is you're in the van. And you get to a gas station where there's a hostage situation. You kick the door down. You're like, get the fuck out of here. It's awesome.

They have this thing. It's this little tablet that folds down from your chest. And that's the mission objectives and everything like that. And I was like, that would be tight if you're hunting with this little flip down thing, deer intelligence. So believe it or not, you can just buy that thing off of Amazon. So I did. And then I went into Revlet. And I completely vibe coded this sick ass app where I have waypoints and everything. It's got a map view. And I can set the waypoints of where the stand is, and yada, yada, yada. And it was the craziest thing, because I had it tap into my cam feed for the cameras. And I have that there, too. It was awesome. And this isn't for anybody. This is just for me. And it's helpful for me.

And it's helpful for me. Three days ago, I went out at 5 in the morning. And it's out in the woods in the dark, like the Boogeyman. And I got lost. I walked past. I must have gone two left or something. And I come up against this like brambles. I'm like, oh, boy.

And I was like, wait a minute. I have this vibe coded app. So I pulled it off. I'm like, oh, OK, sick. And I just walked over here. And then I was fine. I wasn't lost in the woods anymore. Thanks, technology.

All right, now I got to hurry up, because I have two minutes officially left. I'm going to show you my favorite kind of vibe coding, and that's with hardware. So I have this thing at my house. It's this hydroponic garden thing. You can have herbs in the winter. It's sick.

And you can travel, and it takes care of it. It waters it. And it knows the light. And you get text messages about it and everything. It's awesome. I love it.

And I also do outdoor gardening. However, when I travel, I tell my wife, I'm like, hey, you need me a favor? I water that every other day, right? And she's like, yeah. And then she doesn't. And my whole garden dies. It's devastating.

So I'm like, I want to have that for the outdoor stuff, right? So I set up this Wyze cam on a shovel. It's pretty technical shit. So then I go, and I send the picture to OpenAI on a cron job. And it's like, yeah, they're green. It looks good. Like, tight, dude.

So I was like, yeah, I've got to make this more official. So then I started talking to Chad GVD. I'm like, yeah, what would this even be like if we made this official? So it designed this. And I was like, that shit looks sick. I should totally build it, right? And so I'm like, yeah, so how does that work, right? And then it generated this.

It's like the solar controller inside of the enclosure. It's like a battery and a Raspberry Pi and relays and stuff. And I'm like, oh, this is very tight. This is very tight. So I bought all the shit. Remember when I said you have to trust but verify with this stuff? It was so wrong about so many things. I had to go through. I'm like, what is the size of this? What is the size of this battery? I'm like, yeah, no, that's not going to fit in there.

But after getting all the dimensions and power requirements right and everything like that, I went and bought all this stuff. Quite a decent amount of stuff for soil moisture sensors, the motorized ball valve for watering, solar charge controller, all that stuff. So I end up, I have all this stuff on my dining room table. My wife is like, get this off of my table right now. I'm like, all right, I'll build it. I'll put it together. So I put it together. And I started building it. And it looked like the picture. I'm like, this is sick. This is going to be so dope. So then I totally put it together out there in the garden. This is it hooked up, right? You have a Pi. You have the battery. You have a buck converter, some relays, the soil mesh moisture sensors. I haven't written a single line of code yet at all. I didn't even set the Pi up.

It's all just right there. So it's out there with a solar panel all put together, ready to go, but really not functional. Even has a sweet camera and watering. So the first thing I do is I enable SSH on the Pi. And then I SSH in and enable tail scale. If you don't use tail scale, it is the shit. You can SSH into your stuff from wherever; it's very dope. It's my email. Please don't send me photos. You get Termeus, which is sick for your phone or your iPad, allowing you to SSH into stuff all the time and it's amazing. Then you install Cloud Code after SSHing in with Termeus. I made a pin map file with the structure of what pins go where when I put everything together. Then I started VIBECODING intensely, actively working on scripts, temperature metrics, and more. After getting everything to work, I set up a Grafana dashboard for real-time metrics and historical data, which runs off the Pi accessible from anywhere, even when traveling, to monitor the outdoor garden metrics.

And then you install Cloud Code once you SSH in with Termeus. And then all I had to do was I made a little structure here. Because when I put all the stuff together, I wrote down what pins go to what. So I had this pin map. So I put that in a file. And then I started VIBECODING. And when I tell you I was VIBECODING, I wasn't just chilling on a couch with my phone. I was actively drunk in a hot tub all summer long, iterating on this thing.

I'm like, write a script. Give me the aggregate temperature, whatever. Just like VIBECODING beyond what you could possibly imagine. After I got it all working, VIBECODED I need you to set me up a Grafana dashboard where I can get real-time metrics and review historical data and all this stuff. And it set it up and it runs it right off the pie. And because it's on tail scale, I can look at it from anywhere. So I went traveling. And I'm sitting there looking at my sweet, baby, outdoor garden, seeing all of its key metrics.

And what did we end up with, folks? Fresh vegetables. My garden didn't die. We had fresh vegetables. And I did it drunk as hell in a pool. Thank you.

So why don't we take a question about this, the vibes and how it affects these generations of engineers? So you mentioned the recruiting focus on older folks who had struggled uphill both ways in the snow. Did I say it like that? No. OK. Ha ha ha. But what do you think the role of engineers is going to be in five years? Do you think that AI will affect this next generation of coders?

Yeah. It's kind of twofold, right? I think there's some advantage to have always been using it, kind of being AI native, per se. But then also, you think about the impacts of if this blocks off what is traditional entry paths to this career, where do all the new engineers come from? So it's kind of a funny place. I think the role, I can't speak for anybody else how they're using AI today, some less than me, some more than me. But I think in five years, it'll probably look pretty similar, where people will still be building things.

Yeah. And I guess, I mean, I think there's a lot of murmur, murmur about, oh, software engineering isn't even going to be a thing anymore. But as you said, we still need to know what we're building and how it all works and if it all works. So yeah. Well, I guess we'll all find out.

There are also some questions around 3D modeling tools. Have you vibed any 3D models? Excuse me? Yeah, I've tried a little. I've seen people do it. I've seen people use MCPs for that or something. But I tried to have it do some modeling, and it didn't do it. I've seen people get good results. I never got good results. Gotcha. Well, I guess if you have found interesting things, share it with us so we can all learn.

So how long did it take to build? And how much time was on the hardware parts versus the software parts? If we're talking like MVP, I think the software is faster. The hardware is a funny thing, right? In software, we just have node modules. Everything's free, essentially. And ephemeral and make-believe. It's not make-believe. In hardware, you have to wait for it to actually ship to your house, and then it comes. And this thing doesn't fit into this thing. And you're like, what are you kidding me? You have to order a new thing. It's annoying. But that actually came together pretty rapidly.

But that actually came together pretty rapidly. I think I built that in the course of well under a month. And that's with everything, shipping and everything like that. But as far as the software, I feel like it's been a very iterative thing, where I'd look and see how it did. Are these soil moisture sensors as accurate as I want them to be? I added a temperature and humidity sensor. It was a piece of shit, so I got a new one, a different model. And moving forward, now I'm thinking about iterating more, where I'm going to go into KiCAD and do a custom PCB for this. So it can be miniaturized, productized. Things aren't soldered or screwed in anymore. Now we're going to have cables going in. And yeah, it's iteration. Awesome.

Well, and I think we actually have a kind of a follow-up question around iteration and how many iterations you have to go through and what kind of adjustments you make on each iteration. Anything you could elaborate on there? In general? I think the question doesn't specify, but let's say in this project, which you said you were iterating on all summer, how do you decide? OK, now I'm going to tweak this. I mean, sometimes it's because you're not seeing the soil temperature, what was it, soil moisture sensor results you want. Are there other guidelines you use in terms of structuring your iterative development cycle? I think it's just giving a shit. If you give a shit about anything, you're going to iterate, right? Because if you actively care about it, you're going to be like, ah, you know what, maybe not. I think when you stop iterating, it's either done or you stop giving a shit or both. OK. All right, great. I love this. This is, I think, good advice in general when you're trying to build anything. Awesome.

And so I think we have a couple of questions around, yeah. So we spoke a little bit about the future of where this is all going in the industry. Do you think there are some folks that are feeling resentment or maybe have some stigma around AI and vibe coding? And how can we encourage more senior engineers to experiment with AI? I haven't seen that many senior engineers that aren't willing to, in my experience. But I have seen, I don't know if it's resentment, but I think that you have a lot of people who never were in the software field. This is not junior developers. This is dudes who came over who were washed out of crypto. And they're like, I'm building software as a service.

We're fully bootstrapped. And I'm a founder. And I'm like, you're unemployed. You're not a founder. And you're sitting here. And you're talking shit to a bunch of established professionals. So I think there's probably an air of that, especially on Twitter. Where it's like, these dudes are coming in, like, oh, you're washed up. Fucking software is over, yada, yada, yada. And you have a lot of people who are like, what? You're like a cartoon avatar on Twitter. So I think that's probably. But actually, junior engineer or senior engineer at companies, I don't think that rift exists. OK, cool. So maybe there's less stigma than we might feel sometimes by looking at Twitter. Antiquitally.

Awesome. Well, we, I think, just have time for one more quick question. How about what platform did you use to deploy your vibe-coded web apps? And did AI help you do that? Yeah, yeah. Actually, I lean on it more for dev-op-y kind of shit than I even do for coding. So the Mickey Mouse app that went to Linode, which is some server. The Hunting one is on Replicant. It's hosted fully through there. And then the Plant Grafana, that was a crazy spit. Did you see that? Shit was poof. The Plant Grafana was actually hosted right there on the Raspberry Pi. And the Raspberry Pi is hooked up to tail scale. So you go to a tail scale URL, forward slash Grafana, and it just loads up. Awesome, awesome.

Well, thank you so much for the talk and for all the Q&A. And if folks have more questions, there will be a speaker Q&A session in the room just outside after this. So let's give it up one more time for Ken. Thank you so much. Thank you.