The Future of Software: Steve Yegge on How Vibe Coding Changes the Game

Adam Davidson: 00:00:00

Hi, this is Adam Davidson, one of the co founders of FeedForward and your host of the FeedForward podcast. So I have to say I've been a little obsessed with vibe coding lately, and I think a lot of our members are in a similar place.

Curious, maybe a little intimidated, definitely aware that something important is happening today.

I'm talking with Steve Yegi, who is a legendary software engineer and blogger and the co author of the book Vibe Coding, which I read and found genuinely eye opening.

Even as someone who is very much not a developer, Steve has a framework for thinking about what's actually changing, who it's for and what we need to understand right now. Hey, Steve Yegge, it's exciting to talk to you.

I'm a fanboy, so I certainly know you as author of Vibe Coding, which kind of came out at the perfect time.

And it's a great book on both practical steps for people who are a little terrified of vibe coding, but also a lot of inspiration and a lot of thought. I mean, you. The book says it's written for developers. I am very much not a developer. I found it super accessible and easily applicable.

So I'm assuming you, you also wanted me as in the readership.

Steve Yegge: 00:01:09

Well, we went back and forth many, many times about whether to make this book for everybody or for people who have a little bit of coding background. Right? And we, we wound up just to make it easier on ourselves, restricting the audience to people who have written code in the past.

All right, and when I say restricting it, I, I specifically mean, like, we just, we just didn't go the extra mile to like, you know what I mean, like to explain the real foundations, basics of old school programming to people, because then it would have been a 400 page book, right?

And, but I tell you, man, I tell you, the response to this thing has been strongest right on the edge of people who have been out of programming for a while. Right? They did, like your co author. Well, like me, like you. Honestly, I had been out of programming for a while.

I mean, when I wrote the book, I had been programming for 18 months. But it's vibe coding that got me back into it, right?

And we've seen amazing, amazing response from a lot of, you know, senior leaders, people who, who left engineering to go into other roles like product management or whatever. They have a little bit of technical chops, you know, and they realize, oh, wow, I actually have a lot of my technical chops left.

It's just that I didn't keep up with the wave of crap that was Changing as a programmer that you have to deal with the yak shaving to actually get a program written. Right. That was just so much bs, nobody wanted to deal with it anymore. That's why I retired.

And now with the models, that layer is now handled for you. And so it moves you up to this level that all these people with 15, 20, 25 years of experience, they have that level.

There's a new skill set that you need. And. Oh, man. I mean, I could obviously, I could talk all day about it. I wrote a book about it.

Adam Davidson: 00:03:11

Right, right. Yeah. And I want to get to that as well.

I kind of overuse this analogy, but I went through my massive digital disruption experiences being a radio producer, becoming a podcast producer, but at major media like NPR and then seeing, like, the dawn of, like, let's call it vibe radio podcasting. In the early days, I would have friends call me up and say, how do I create a podcast?

And I'm like, the way I create a podcast is I hire five people and they create a podcast. So.

And so entering a world where, like, you buy a 200 mic and some cheap software and you just start talking, it took me a while to understand the logic of that world. And it's interesting how every field, I guess, like, in journalism, certainly you're an amazing journalist.

You've covered wars, you've covered, you know, so how are you rewarded? We're not going to let you do journalism or you don't want to do journalism anymore.

We're going to make you a manager of other people doing journalism. But with coding, you actually can kind of have create your own team of coders every second with AI tools.

Steve Yegge: 00:04:17

Yeah.

I mean, look, that point that you brought up about podcasting becoming accessible to everybody, you don't need to have a bunch of money to do it anymore. And a team of experts, that's happening to software development, and that's really freaking a lot of people out.

You can even plot out which demographic is the most freaked out about it. It's like it's engineers with 15 years of experience. That's the sweet SP of people who are revolting the most against this.

Adam Davidson: 00:04:48

Interesting. Like, more they like it and less they like it, but that.

Steve Yegge: 00:04:52

Yep, that's right. That's right. That's right. So you.

Adam Davidson: 00:04:54

Oh, that's really interesting.

Steve Yegge: 00:04:56

Yeah.

The more you have invested in your current career, which peaks at around 15 years, once you get to 20 years, you're like, yeah, I've done all this stuff. Right? But like, yeah, people in the 10 to 15 year mark are really resisting. Yeah.

And but what's happening is man, everybody else is picking it up, right? And it's like, like digital photography used to be photo. I mean photography used to be photographers only and they were really snooty about it.

And when everybody got a camera and everybody was able to make, it just became a non. The question wasn't even a question anymore. It wasn't, oh, are you allowed to is digital better than the old school?

Blah blah blah, all those arguments, it was all just noise and it all vanished. And now everybody takes a gazillion pictures and it's this sharing content sort of exercise that everybody can now participate in.

Software is happening. I mean the exact same thing is happening in software, okay?

Software is going through the exact same transformation, which is weird because software, like photography before used to be this thing that only experts could produce and now anybody's going to be able to produce software and it's going to be a weird world where we're all going to be like, really me? I can.

Oh wow, here, here's some software I made and we're going to share it with our friends and family and realize that wow, it's just like creating any other kind of content. Right.

Adam Davidson: 00:06:10

And it doesn't mean there's no coders. And we'll get to that. Like it might mean there are more coders.

Steve Yegge: 00:06:14

I just got off saying are there no photographers professionally now that everyone has a camera? No, it doesn't work like that.

Adam Davidson: 00:06:22

Yeah, my guess is it's not great for mediocre photographers, but it's actually extra good for really skilled photographers because all of us are not just picture consumers, we're picture takers. And we can really now viscerally tell like oh, that thing's doing something I don't know how to do. And that's really cool.

And I'm already experiencing that with coding. So in our membership we have some people who are extremely technical, but we have some people who have not really cracked the seal on.

I think a lot of people, probably most people. I just spoke at a big company and they had done an internal survey and 1% of people in the company had done more than just chat with AI.

And so I want to.

Steve Yegge: 00:07:06

Well, that sounds right. That sounds right.

Adam Davidson: 00:07:07

Yeah.

Steve Yegge: 00:07:08

One percent sounds about right. It's insane.

Adam Davidson: 00:07:10

Here's a few things I've noticed and I just want to kind of talk through.

Like when we talk about coding with AI, maybe we can define the terms like you can do within say Claude Code or AI Studio at Gemini or Codex or ChatGPT. Like, it can do a lot for you that we could call coding. Right.

You could say, spin me up a web that does this and this, or even a web app that does this, this and this. And it'll just kind of do a bunch of stuff.

You can use one of the consumer tools like Lovable or Bolt or there's a bunch of these where you just kind of write a prompt and it'll generate whatever it is you want. A website, an app. So that's one version.

I think we're talking about, like the next step where you're actually, like, you're not just seeing it as I write a prompt and then out comes an app. But you're actually in the mix.

You might not actually understand the code, but you're in the mix actually thinking through, like, how are we structuring this tool? What are the choices we're going to make?

How does the front end, the thing that the user sees, relate to the backend, the kind of database and the tooling, making it run? Is that the universe we're talking about when we step up the ladder of vibe coding?

Steve Yegge: 00:08:27

Yeah. And you're going to be able to make more and more sophisticated software without understanding the software or seeing the code.

And you're going to be having a discussion with the AI about how you want it to work. And it's going to be really fiddly because software is really fiddly.

You're going to be like, well, look when I've logged in, but I've wired this resource over here, and then I go to this screen, like it's supposed to do this, right? And that conversation, it's gonna start to bog down after a while.

Like, you're gonna be able to accomplish a lot and you're going to be really pleased with it. And maybe it'll actually solve your problem. But at some point, your ambition will exceed your ability.

Just like if you want to be a photographer, you know, you're going to try to take some pictures that are sort of outside of your comfort zone and you're going to stretch yourself and software is going to be the same way.

And so that's why, I mean, that's where you gotta understand software can get bigger and bigger and bigger until you've got a piece of software that's the size of the Death Star, Right? That's an. That's an. That's an enterprise, Right. Remember, remember the under construction Death Star in Empire Strikes Back?

Adam Davidson: 00:09:35

Yeah.

Steve Yegge: 00:09:36

That's an enterprise project, right? There, right? Like they're trying to build something as big as the Death Star. It's about half finished, right.

It's going to take years and years and years.

So just because you can build software on your desktop to replace some of the applications that Apple made that you don't like, which people are doing, doesn't mean that you can build a Death Star. That's going to take professional engineers and probably a hundred of them, right? So, like, software comes in all sizes.

And what's happening is professional engineers are getting pushed up into building larger software because regular people are now programmers and they can build smaller software. You see what I'm saying? So if you have training as an engineer, it just increases the amount of ambition that you can have. Right, Right.

Adam Davidson: 00:10:20

And a lot of my friends, I mean, just to, you know, when I was at npr, there was like, I don't know, four or five jobs in the country that were like, I run a bunch of radio shows and at least like public radio, long form. And now a lot of my friends run, you know, massive operations. They make a lot more money, they have a lot more shows under their belt.

Steve Yegge: 00:10:46

So how come I'm not talking to them?

Adam Davidson: 00:10:49

You probably should, honestly. Okay, so.

Steve Yegge: 00:10:55

No, but your point is important, right? Which is that it's commoditizing and democratizing content creation, and that's going to accelerate because software is just content, right? In a way.

And so we're going to see an explosion of interesting content much, much more to select from than we've had before. Which is good, right? Because it always feels like there's not enough good content out there. So I don't know. Right, Agreed.

Adam Davidson: 00:11:23

Yeah.

I remember thinking, I remember in the 90s, I was talking to my girlfriend at the time who was a radio producer and thinking, I just want an hour of good TV a day and an hour of good audio a day. And why can't an economy the size of the United States or the English language world produce that? And now.

Steve Yegge: 00:11:43

And the answer is, I'll tell you why. It's because we're getting pickier. Like when I was a kid and there were seven stations on tv, man, you picked one and you watched it.

Adam Davidson: 00:11:51

Right?

Steve Yegge: 00:11:52

That's true.

Adam Davidson: 00:11:52

Although I will say that now it's the opposite problem, right? It's. There's definitely over an hour. There's. I could probably spend 24 hours a day listening to pretty compelling audio. I could definitely do that.

There's millions of podcasts. It's. Now how do I find it? How do I Sort through it, but we'll, we'll get to that in a minute.

Steve Yegge: 00:12:11

So.

Adam Davidson: 00:12:11

All right.

So one thing I want to encourage our non technical members because this is something I've come up against with like my wife with some friends, is there's a few hurdles that can feel insurmountable to get into this world. But I want to argue they're not hurdles.

The analogy in my mind is 20 years ago or something, I took a French cooking class and I was very intimidated by cooking. I didn't understand cooking. It was six three hour classes. They were really fun, we ate really well.

I am not a master cook, but just those six hours, it just pushed me through to where I'm never going to beat a master chef. I'm never going to cook better than a master chef and I'm never going to even cook better than my wife or almost anyone who spent.

But I now have the ability I can order out, I can give the prompt of I want this meal and someone else will make it for me.

Or I can actually think like oh, I kind of want a steak, but I like it, you know, oh, I want more crunch or I want some acid or whatever and I'm competent enough that I can make what I want pretty good. I can't make it amazing and I can, you know, not embarrass myself. If some friends come over, they could definitely get a better meal somewhere else.

But you know, and I think there's like the six evening French cooking class equivalent of coding that I wanted to talk about the command line interface and IDEs and how once you break through that which is at first can look like I can never ever. What is that? I don't even understand what that is. I'll never use it. I think that opens up a whole world. So maybe just walk me through what those are.

Am I right? That that's kind of just push yourself through, just learn how to use those and a whole world opens up.

Steve Yegge: 00:14:09

We have like one pretty good data point to support this, which is that Gene and I gave a workshop at the ETLS conference in September in Vegas and we had about 60.

Adam Davidson: 00:14:20

I'm sorry, what's the ETLS conference?

Steve Yegge: 00:14:23

The Enterprise Tech Leadership Summit, which was. It's a summit of, you know, lots of friends of Gene, basically which is everyone in the industry.

And we held this workshop, a vibe coding workshop for engineering leaders. Yeah. And it was just absolutely magical. It was unbelievable. We vibe coded the whole thing together.

About 45 minutes before it started we realized, oh no, people aren't Going to be able to NPM install cloud code or whatever. And so we need to set them up with accounts on some cloud box that they can log into so they can run the tool for our workshop.

And so we, you know, we asked Claude and we vibe coded the whole thing up and it came together and it was like the whole, the whole, the whole show was like seat of its pants. We didn't know how it was going to be received. And yet, dude, people loved it.

They loved it and they wouldn't shut up, they wouldn't stop talking about it. It was the highlight of the conference. These people, they hadn't coded in 15 years, 20 years.

For a long time, whatever they're, you know, VPs, CEOs, CTOs, whatever. They sat down and we talked to them for 20 minutes about just how you argue with AI to get code out of it, right?

And then we, we cut them loose with their accounts and we said, just, just start playing with it.

We gave them some data sets like the Titanic Passenger list, just some cool data sets that they could like, you know, do stuff with visualizations or applications or whatever. And every single one of them, every single person within a half an hour was producing software. Okay.

And by the end of it, 70% of them had something like, showable that they could, that they could show. And all of them had had fun building things. And so, yeah, to answer your question, is there a, you know, a six, three hour course?

I mean, there's a shorter one than that. You, you get into a room. We spent four hours and by the end of it we had a bunch of vibe coders.

Adam Davidson: 00:16:20

Yeah, that's awesome. And maybe we should get you that for our membership. I'm sure they would love it. Let's talk about that.

This is like whatever computer you use, a Mac, a PC, Linux, if you're using Linux, you probably know this already, but you open a terminal, they all have a terminal. I don't know, Windows, they call it a powershell. But you can also just ask Claude or ChatGPT how do I open a terminal on this computer?

And it'll tell you.

And then you just, I mean, in fact, you can just ask Claude, Gemini or ChatGPT how to do all of this, but basically you go to a terminal, you install Claude code, or you could install Codex, the ChatGPT equivalent, and then you launch it. And then it looks a little weird to me. It looks like a computer from the 70s.

It's just text or if there's an image, it's a very, very Old school ASCII image. But you just talk in a way. You just talk to it like you would talk to a chat interface. But it, it doesn't have that narrow.

I mean, the most annoying thing with the chat interface is you just inevitably hit the context window and it just shuts off. You can't do anything more. But also it's not creating files for you alongside whatever you're saying.

This will create little files that contain the programming you need. So once you get past that little step, that's where imagination can you know you at a first pass. You can just do what you do in cloud code.

You could say, my son's really into mountain climbing and geography.

Can you create a game for a 14 year old kid who wants to climb the highest peaks around the world and make it a fun interactive game that helps him learn French conjugation? I might actually do that. My son's struggling with French conjugation. He really wants to mountain climb. Maybe I could try that.

Whatever it is, whatever you do.

Steve Yegge: 00:18:24

The House of Etre. Yes. French conjugation is pretty crazy. Yeah, right?

You can build that application and I mean, like, honestly, like to get started, you could even start simpler than that. The thing about cloud code and sourcegraph Amp and Codex, you know, and Gemini Cli is they're concierge assistants.

They're helpers that can help you with anything on your computer.

And so if you need software installed, or if you have software that's misconfigured, or if you need things upgraded, or if you just want to move things around on your computer and clean stuff up, whatever, you can actually just tell your agent to do it.

And when you get into that mode of have the agent do everything for you, you can reorganize directories, clean things up, find pictures for you, whatever. Then, right?

Then it becomes easier for you to get into that mindset that you need to be in to say, well, I need to do something that maybe we don't have the software for, but maybe we could build something.

Adam Davidson: 00:19:27

Yeah, that's a really good point because it doesn't have to be software. I mean, what have I done with cloud code? I had a downloads folder that had a ton of files that was too.

So I just said, go through this, put them into folders and also give me suggestions of what I can delete. And it did a great job.

Steve Yegge: 00:19:42

So, I mean, you know, you start that way, you should start by saying, look, this is my computer concierge and it is also my coach, right? I mean, like I'M absolutely happy to have the thing coach me.

I'll be like, we'll be going along and talking about the architecture of the system, and it'll say something that I don't understand. And I'll be like, can you tell me more about this? Because I don't. That's not familiar to me.

And I don't know if that's like an interaction that a lot of engineers feel super comfortable having because of their pride. Right. They're like, whoa, I can't admit that. I don't know what it's telling me. Right. You know, you have to like.

But, like, I mean, do you know everything in Wikipedia? No. You know, are you. Does it hurt your pride to read a Wikipedia page that you've never heard of before? It shouldn't.

And so it shouldn't hurt your pride to have the LLM tell you something that it's never seen before, because it's really just sort of a fancier version of Wikipedia. So once you get over that hump, man, it's at that point, it is like a person. A person with interesting quirks. Right.

That you need to, like, kind of work. Work with. But, yeah, you're having a conversation and you're just telling it you're not happy until you are happy.

Adam Davidson: 00:20:55

Right.

Steve Yegge: 00:20:56

And again, that's not a job a lot of people want. A lot of people don't want to just order a computer around to produce the result that they're looking for. Yeah.

Adam Davidson: 00:21:06

Right.

Steve Yegge: 00:21:08

And so I don't know, man. We're in a really weird sort of transition.

Adam Davidson: 00:21:11

Although it is a kind of. So that first level, before we're writing any code, we're just embedding it into our computer.

We're taking it out of chat, putting it into our computer. It is crazy. Like, the other day I had this super annoying problem where the computer was when I talked. The mic showed the green light moved.

It could hear me, but my transcription software couldn't hear what the mic was saying. And I spent like one of these annoying two days Googling, doing different things, trying to troubleshoot it. I couldn't figure it out.

For some reason, I hadn't thought of this. But then I just fired up quad code. I said, this is what's happening. Figure it out. I have no idea how it figured it out.

I don't know what the problem was. I don't know what the solution was. It fixed it.

Steve Yegge: 00:21:51

Yeah. No, seriously, you shouldn't be troubleshooting your own software installations anymore.

My obs was misconfigured I'm just like the scene switcher plugin wasn't loading or something. That's never something I would do myself again. Right. I wouldn't do that myself anymore.

I'd go to Claude Code and I'd be like, or in my case, source graph amp. But since we have to say Claude code because it's a class, it's like Coke is soft rigs.

Adam Davidson: 00:22:21

Right.

Steve Yegge: 00:22:22

It's because fewer than 1% of the world's developers are actually using these tools. You see what I'm saying? So agentic coding assistants. But they're agentic assistants.

And so your first thing that you should be doing is having them help you with all of the software on your computer. Just configuration.

Adam Davidson: 00:22:39

Yeah.

Steve Yegge: 00:22:40

And I think that's how you get your feet into this world.

Adam Davidson: 00:22:42

A buddy of mine has a folder with all the manuals for all the stuff in his house, like the refrigerator, the boiler. And just when he hears a weird sound in his house when something's not working, he just asks, hey, yeah, what's going on? Figure it out.

And it both can just read that. But it has figured out implications of his.

Like, my electric bill was too high this year and it was able to figure out something that probably even if you hired an electrician would have struggled to figure out. Yeah. Another one I love is a buddy of mine. He uses granola to take meeting notes for every meeting and transcripts and he keeps it all in a folder.

And then he told Claude code, read all of these and tell me where I'm avoiding conflict. Because he felt like he was being too passive in business meetings.

And so it identified and he thought it did a pretty good job of being like, oh, in this meeting they were pushing back on this and you just let them. And you could have done this, this and this. Tons of people are using it for sales advice, blah, blah, blah.

All right, so getting in there, having fun, starting with just whatever's on your mind.

Like, you know, download a bunch of recipes you like and, and ask it like, what do you think are recipes that if I like these, what else might I like? Or what's a kind of cuisine I haven't tried that I might like or whatever you can think of anything in the world. Here's my 30 favorite movies.

Give me a list of 30 others that I might want to watch. And you know, anyway, so there's all that stuff.

But then getting into code, I think what's happening to me is that as a non coder is my mental frame has shifted from software is a thing. It's like I buy Microsoft Word and that is Contained and that's a thing. Or I buy. I use Zoom and that's like, it's got all these parameters.

But I've not coded anything approaching the complexity of one part of Zoom.

But I've coded enough now that I realize software is just a bunch of human decisions, often like that to overcome problems or, you know, and so now I. Whatever I'm using, I mean, right now we're talking with Riverside. I don't. It's not like, oh, that's a finished product. And that's just what it is.

I'm thinking of all the choices they made. What choices would I make differently? And I feel like it just re engineers how I understand the world. It makes me feel like I have more agency.

Steve Yegge: 00:25:07

Yeah, software's moved for you from the flow, from the integer world. The software is. Or it isn't the floating point world. Floating point, right.

Where it's this accumulation of many, many, many small decisions, many of which now in the future will be able to be changed. I think software is going to become more malleable as the more successful software is going to be.

I mean, like here, here it is again, like full, you know, 12 years later or whatever platforms. Actually 14 years since my platforms ran. But every piece of software that is structured as a platform is now going to.

Going to win in the marketplace. Because why?

Because the platform is going to give the AI access to it, which means users will be able to change it and customize it to their own personal tastes and needs. Right. And so that whole idea of software being the shrink wrap thing that you can't change anymore is going to go away.

All the software that you have is going to be changeable or it will be replaced by software that you can change. Right? Because you'll write your own version of it. Like Gene Kim, he just wrote his own, I think, Preview app, you know, Preview on the Mac.

You know, it's. I don't know if you're a Mac user, but Preview is how you.

Adam Davidson: 00:26:20

I am a Mac user. I hate Preview. So limited.

Steve Yegge: 00:26:22

Like, like if you're using Preview, you like, you're stuck with it. It's just, it works the way it works, right? But Gene, like, was like, screw this, man.

I need a better PDF viewer that allows me to like arrange them in this exact form. And da da, da, da. He didn't like Preview, so he Vibe coded up a replacement for it in a couple hours, right?

Which to me is just absolutely mind boggling Right. Because I mean, like you, I think of the apps on my Mac as these. Integral. Right. It's a whole app and you get it and you're stuck with it.

And Apple made the decisions and that's the end of it. Right.

Adam Davidson: 00:27:01

And if you're lucky, you can choose from sheets, Excel and whatever.

Steve Yegge: 00:27:06

Yeah, right. And they probably don't have much customization, but Gene was like, screw it, I don't like your app. I'm gonna write a new one.

And he wrote a new app and that doesn't even use Preview anymore. Right.

And it's like, well, damn, if you can replicate the work of a three to four person team inside of a big corporation, then you don't need to rely on their software anymore. Right. So that whole class of smaller consumer apps is now. It's malleable, you can make your own.

And so this is why I'm saying Preview, if Preview wants to be successful, it's going to have to change so that AIs can modify it. Right. For you. Right, Right.

Adam Davidson: 00:27:39

Absolutely. Yeah.

Steve Yegge: 00:27:42

So anyway, I think platforms, platforms are going to resurface as the new sort of accessibility layer.

Adam Davidson: 00:27:48

I keep thinking like a lot of software has a thing I like, but not. But then a bunch of things I don't like. So why should, why not just have the thing I like and then I can access it.

However, like, I can create my own interface or I could create my own. Yeah. Way of. I mean, I, I really hate Preview. I think it's, you know, I.

But then I bought like wondershare, which brings its own problems like PDF viewers. Don't get me started on PDF viewers.

All right, so we've talked about and something we really encourage at feedforward is just vibe code on the weekend. There should be the lowest bar in the world towards trying something. I'll just encourage you. Get over the intimidation of the cli.

The command line interface.

It looks weird, but once you realize, oh, there's just one line I got to worry about and then it says some stuff and then it prints out a whole bunch of stuff I can't understand with lots of like weird code language. Just ignore that. Just read the English sentences, you'll be fine. Certainly in the beginning.

Now I want to describe the world in which, you know, large company XYZ traditionally had developers somewhere in a corner or somewhere in another building. And it had subject matter experts or executives. And for them to interact was probably like a big event.

Like someone in their career who's not a coder could remember, like the one or two or three times they had a big thing with a coder, with a coding team. And now we're in a world, you know, Walmart and some other companies have already sort of pushed developers into departments.

So there can be, you know, the marketing department or the shipping and logistics department now. Now there's a software developer sitting alongside them. Overall, I find that exciting. I also find it really challenging to how work is done.

And I can imagine as a developer like a little bit of a, like I'm not sure I want every single person in this company creating code. That seems like a lot of code. A lot of it'll be crappy.

Just walk me through how you picture we're in a rocky period maybe, but like a year, four years from now, whatever, when this is kind of stable, what do you imagine that interaction between a near ubiquitous vibe coding alongside pro developers?

Steve Yegge: 00:30:09

I mean, I see it heading exactly in the direction you just outlined, which is it's going to bring the engineers closer to the business. Okay, we're already seeing it. The AI can provide like a translation layer between the two.

Business people can speak a little engineer now and engineers can speak a little business now. Because you can go to the AI and you can say, I'm about to have a conversation with this business owner.

They're the country head for Malaysia and I really don't know anything about Malaysia, but I need to talk to them about some software we need to deliver for them. Help me prep, right?

And so when you go in to have a conversation with them, it's no longer two different universes that have this huge impedance mismatch, right? There's some common ground which is new.

And I can tell you, having worked at GRAB in Southeast Asia where I had to work with business owners like the country heads of Indonesia and Thailand and Malaysia, having that translation layer is life or death. Because I would need to roll stuff out to them and there was just no way to do it.

It was an eight month conversation and the business was perpetually angry with engineering. They hated each other. And it was funny, it was institutional. Like we would tell them, we'd say, why are you so mad all the time?

And they say, oh, you need to be more thick skinned about it. That's just how we are. We need someone to blame. It's like, okay, right? And so like now with AI, it's going to be completely turned on its head.

You're going to go to the business owner, you're going to have a shared prototype that the AI is creating that you're making together. And the business owner is like, no, our customers don't understand this, this, this. We need to change it to blah, blah.

And you'd be like, okay, do you like this or this? You can make options, you know, and whatever. And it becomes this, this, this, this conversation rather than this confrontation. And.

And honestly, I'm looking forward to it because, boy, I've had a lot of confrontations with business people.

Adam Davidson: 00:31:56

I bet. I mean it. And I'm gonna just keep talking about podcasting, because why not? Well, I'll talk about radio. So I.

When I entered radio in the early 90s, you know, there was the engineer, the soft. The audio engineer. And a lot of tasks were just done by that audio engineer. And then we got desktop audio editing. You know, that started.

really ubiquitous till, like,: 2009

And the audio engineer union, their response was to try and restrict what people could do. Try and, you know, it was the opposite of, like, let's integrate. Let's. Let's be a.

You know, I remember talking to one of the union leaders and saying, hey, why don't you do it? Where you just are really helpful as, you know, because more people. And she literally was like, huh, Wait, how would that work?

Like, it was just a mental frame that made no sense. And eventually npr, of all places, broke. Broke that union and. And desktop audio editing became ubiquitous, but we missed the, like, expert layer.

We weren't able. You know, it would have been better if it was more, you know, what became it became.

There were audio engineer things, and that involved, like, filling out forms and, like, really strict rules. And then there was the stuff I could do on my own. And it was very hard to gather that expertise to go back and forth.

The engineers felt we were threatening them and trying to take their job away. Our bosses didn't want us using them too much because it would strengthen the union. I mean, it was weird. It was weird. We want to avoid that.

One thing that I find really cool is I have an idea for a thing, you know, and I say it in layman speak, like, oh, it'd be cool if there's a website where I could just dump some documents and it'll tell me how to write them better. Say something like that. And I have a point of view about how to do that.

When I start vibe coding, I'm sort of forced to ask engineering kind of questions that I wouldn't like I did something like that recently. And it's like, wait, do I want it to be? Am I. What am I dumping in? Is it. You know, I have to decide, is it any document, PDF, word, da, da, da.

Is it a markdown file? Is it a text box? And then what happens? Do I press a button? Does. Is it. Does it automatically work? What do I trigger? Do it. Does it.

Do I trigger one output? Just generic advice? Or do I have three buttons, like short advice, medium advice, whatever it is. And. And it force.

So this is where I want to get into. Because you built this tool beads that I use because.

Because I feel like I'm in a sort of bull in a China shop way, like clumsily stomping around stuff. You spent decades. My journey was, you know, initially chat. Then I got into the consumer tools lovable bolt. And that was really fun.

Like, create a web app that'll help my son with this. Create a. My own personal blog. And it does a pretty good job. Then I pushed through and I started doing more, you know, actual software.

And I realized a bunch of things that I assume you learn. Like, first month of, like, software training. Like, I hadn't done the most basic stuff of like, what am I building? Why am I building it?

Who am I building it for? What's the. What would good look like? So that was a whole journey is realizing, like, there's a lot of questions to ask at first.

In some ways, it doesn't matter with vibe coding, at least in the beginning, because you can just say it in a bad way and see the tool, and because you can iterate. You can do a crappy version, say, wow, that was crappy, and then redo it. But if once you're ready to go, it's helpful to do that. You know what?

I've learned all these phrases, user stories and specifications and constitutions and all this stuff.

Steve Yegge: 00:35:54

Yeah, it's kind of. It's teaching you. It's teaching you software engineering as you do it. You're learning. You're learning the right.

Just like if you were in a kitchen learning how to cook with master chefs, you'd be gradually picking up how to cook. It's like totally a learnable skill.

Adam Davidson: 00:36:09

Yeah, it's a learn. And there are these.

Then the thing that I that hit me, and this is so obvious now, but I would just kind of vibe code I would hurdle in whatever direction. I'd say, hey, make it so that the flashcards on my flashcard app flip this way.

And then I'd be like, oh, Wait, the admin screen should allow me to resort by category, not just alphabetically. Oh, wait. I also don't like the colors. Make it look like this or. Oh, no, I got an idea. The user should pick the colors. Oh, wait, I want to.

Oh, I have a bug now and I fix. And so I would just mix up all these processes, and the AI is happy to go wherever you take it. It'll get.

It'll just drop everything it's thinking about and just hurdle like a kid with ADHD and whatever direction is new. And so I faced this major problem of just. I wasn't completing anything and I was building, like, these shaggy.

And so Matt Bean told me about beads, which is the tool you built, which just kind of forces the AI to stand. Like, I can actually be pretty adhd. So maybe talk about bead and talk in generally about the architecture of the process itself that one needs.

Steve Yegge: 00:37:23

Yeah, I mean, like, it's just, you know, 50 first dates, you know, Adam Sandler, Drew Barrymore. Right. That's how this is.

Adam Davidson: 00:37:31

Every morning, she wakes up having no idea who she is, who he is. And so she creates this little video every morning that tells her, this is who you are. I'm your husband.

Steve Yegge: 00:37:40

That's right. She wakes up. She wakes up every morning and she has to watch the video. Right.

So that's your context that you give every time you start an agent, which is about every five, 10 minutes, because their context window fills up and then they fall over dead. You have to restart them. You need to refresh their memory on what the context is, where they are. It's exactly 51st days. You have to land them.

And the faster you can land them on the task, the better. And then it's like breathing. It's this rhythm. It's like you make a friend. It's like, what? Fight Club, Single serving friend.

You remember the whole single serving friend thing?

Adam Davidson: 00:38:14

Yeah.

Steve Yegge: 00:38:15

On the airplane, you make a single serving friend out of this agent. It understands your problem space really well. By the time it's finished its task and it's getting ready to go. Right.

And then it completely wipes its memory again. Start over. Who are you? Hi. And what are we still trying to do? It's so frustrating. Every five minutes you get a new worker.

It's like you're running a restaurant and you get a new employee every 10 minutes.

Adam Davidson: 00:38:38

It really is, Matt. I mean, and we should just note that while I love Vibe, it's so fun, I get addicted. I can spend 14 hours on it.

And just doing it I would guess 90% of the time, I'm like, screaming at the thing about how terrible.

Steve Yegge: 00:38:57

Yes. Yeah, yeah, absolutely.

It'll give you a heart attack because it'll do the right thing nine times in a row, and on the 10th time, it'll do the stupidest thing ever. And you're like, what's. Well. And. Yeah, that's why we wrote the book. Right. You know, this is to teach you the. The skills that you need.

But I mean, basically, yeah, you're just having a discussion with it, and it goes wrong a lot. But. But I tell you, I mean, like, you. You. You get this. You get it down, and you can build with it. And it, like, you say it's incredibly addictive.

It's incredibly addictive, man. I mean, like, so, I mean, like, if you can get over the hump the. Of getting started with it. Yeah. You know, admittedly. Okay, so back to the amnesia.

All right. The first thing you're going to notice is that these things, they. They're like a little kid soccer team.

Like, you say, they'll chase whatever you tell them to chase. They'll go off and they lose their focus. And so they'll say they're done.

They've been trained to tell you they're done and present a very nice completed list of things that they did. But unfortunately, they. They leave a lot of crap around. Yeah, they. They leave, like, it's like you hired some house cleaners.

They're going to come in, they're going to do 60 to 80% of the job. They're going to present you and with a receipt that says all the things that they did.

And now you have to go back and you have to say, look, you didn't do the bathroom. You need to redo the carpet. You accidentally left the couch upside down. You know, you have to go through all the things that they did wrong in.

In a cycle until they're finished. And this is weird for people because they don't. They don't want an advice.

Adam Davidson: 00:40:28

Right, right, right. Exactly. And for some reason, you.

You completely remodeled my bathroom and turned it into a sauna when I never asked you to do that, but yeah, or you turn. You turn my plumbing into my electric and my electric into my plumbing, and I don't know why.

Steve Yegge: 00:40:42

You know, what happens is people try it and then they go, it remodeled my house and turned it into a disaster, and it's like, well, no, actually, you turned your house into a disaster by not having it finish the remodel and by having it finished Having it finished means you need to provide the drive, which. Which effectively means Taft Decomposition. Okay? You can't say, remodel my house.

It's just going to go and paint everything blue, and you're going to be like, what the hell was that? Right?

You have to say, I need you to do the curtains in the master bedroom first, and we need to hang them eight feet off the ground, and they need to be this color, right? You need to go, right, you need to break your tab.

Everything needs to be broken down into the smallest possible thing that the AI could actually do that would move you, meaningly, forward, not clean the room. It's pick that shoe up, okay? And once you get into that mindset of that's how you're remodeling your house, and it becomes this long process.

It could take hours, days, weeks of vibe coding to get this to the software of the state that you need it to be in. That's. That's when you have your real breakthrough, right? Because now you. You don't trust it.

You understand that it's a tool that's always going to be 60% and to 80% accurate.

And you're going to do successive refinement to get, like Newton's method, you can approximate a root to arbitrary precision just by these approximations, by making them slightly better each time. That's exactly what you're doing with vibe coding.

It's an approximation that's blurry, and then it gets clearer and sharper as you refine it and say, no, no, no, that's not what I wanted. And it'll go off the rails and it'll break something that was already working. And you're like, ah, stop it.

And that's why verification is so important. You have to have tests that will fail, and you have to tell the thing, make sure the tests don't fail. It's a constant struggle.

It's a struggle to get these things to work.

Adam Davidson: 00:42:30

But your tool beads, and there are other. There's spec kit, there's others of these tools.

Steve Yegge: 00:42:36

Okay? So sorry. Beads is different from all the other ones because the other ones are all for planning. And beads punts on planning.

You can use beads for planning, but it's not really for planning. Beads is for execution. Beads is for saying, I have a plan. I have spec'd out. File an issue to pick up the shoes.

File an issue to dust all of the, you know, furniture. File an issue to change the curtains, right? And what you do is you file a bunch of issues. It's an Issue tracker. And now nothing is ever forgotten.

It's video cassettes all over the house for your AI to come look at when they're confused. Right? And so with beads.

With beads, it's basically like you have taken your plan and you've made a bunch of little teeny tasks with numbers, like, with actual issue IDs that you can refer to by name. So pick up the shoe is issue, you know, BD375. Right. And you can tell your agent now, you can name pieces of work. You can say, let's talk about BD672.

Like, this was like.

And it'll look it up, and it'll be like, oh, this is the one where, you know, you had a problem where one user couldn't get their installation working. Right, Right. And so it gives you memory. It's memory. It's literally you've given a memory module to the AI because. And why didn't this work before?

Everybody was sticking stuff in markdown files, okay? And the markdown files would proliferate.

They would accumulate until you had so many of them that the AI would get amnesia from the markdown files that it was supposed to be using as its memory. In other words, think about, like, memento.

If that dude wrote messages to himself all over his body, like, every square inch of his body was tattooed with reminders, he wouldn't be able to find his way. Right. He'd be like, okay, where was that phone number again?

Adam Davidson: 00:44:24

Right, Right.

Steve Yegge: 00:44:25

You can only have a.

Adam Davidson: 00:44:26

You don't have a. And you might have, like. Right. You might have contradictory ones as well.

Steve Yegge: 00:44:30

Yeah, exactly. And that's what happens with the markdown files is they start to. They rot. They start to contradict each other.

Software is this living thing, and it sort of moves forward and things, you know, they get fixed and then they rot a little bit. And that. That actually reflects in the markdown files.

And then all of a sudden, you're dealing with an agent that seems smart but actually has schizophrenia, and you don't know it until five minutes into the conversation when you realize they were working on something that you canceled two weeks ago. Right, Right. But the markdown file is still there. And so, like, what. Basically what Beads did was. It was like, all right, everyone shut up.

Everyone shut up. Right? Like, seriously, that's how I got started with beads. I was like, stop.

I want you to put every single thing that we're going to do together, all of it, into an issue tracker. Just screw it. All right?

And at first we were thinking GitHub issues or Jira, but they didn't do what it needed, which was provide a graph of all the work so that it could actually, like, reason about dependencies and what's next and doing things in the right order. Beads lets it do that. And so I was like, okay. And basically the AI designed it and it's been this. Man beads is not like a big deal.

It's not a big project. It's a little bit of code. It's a drop in that you use with your agent that gives it memory. Right.

So you've got this employee that's always losing their way. It's a little harness that you strap on. I mean, it keeps them focused. Right. But man beads is like a tiny, tiny piece of the thing that I'm building.

Adam Davidson: 00:46:03

Maybe we could have an off non recorded chat about all of this.

Steve Yegge: 00:46:06

Absolutely.

Adam Davidson: 00:46:07

We'll find some time next week. But you made me realize I could use spec kit for the planning or something like that, and then beads for the issue tracking.

Steve Yegge: 00:46:15

Yeah, because it produces a plan and then you say file all these in beads. It's great.

Adam Davidson: 00:46:19

This was awesome. Steve, thank you so much. Thank you.

Steve Yegge: 00:46:22

Yeah, my pleasure. Happy to come back and chat about. Chat about this stuff again anytime.

Adam Davidson: 00:46:27

And let's do some vibe coding with our members.

Steve Yegge: 00:46:29

This sounds great.

Adam Davidson: 00:46:30

All right, Steve. Yegi, you can find his book vibe coding wherever you get books. Please let us know what you'd like to hear more of on this podcast.

Are there people you'd love us to talk with? Let us know. As always, thanks for listening and see you in the discord.