Latent Space: The AI Engineer Podcast - Steve Yegge's Vibe Coding Manifesto: Why Claude Code Isn't It & What Comes After the IDE

Starting point is 00:00:03 We are here live at AI Engineer Summit with Steve Yogi, the legendary Steve Yogi of Stevie's tech talks, Steve's platforms, brands, and most recently, Sourcegraph and AMP. Welcome. And most recently, VibeCoding. That's right, the VibeCoding book. So this is the big vibe coding discussion. In the pre-chat, we were discussing the intersection of vibe coding and AI engineering. So we've got the kind of movement leaders of both sides here. How do you see it?

Starting point is 00:00:28 It's absolutely a movement, right? You've got to get people behind it. I said at the end of my talk today that there's a huge backlash, and the backlash is only just brewing now. So you and I are pushing forward, right, on these waves of, you know, AI engineering is about building AI-enabled applications and being in AI, and vibe coding is about abandoning the old ways of producing software and embracing the new ways, right?

Starting point is 00:00:54 And both of these are making people pretty mad, right? I don't, I think there's, They're mad if their identity is tied to the way that they work today with no changes, no room for changes. Yeah, so I'll start with my first hot take. Okay, let's go. There is a demographic that is the most affected by that. Their identity is the most tied up with the way that they work. Okay.

Starting point is 00:01:19 It's not junior engineers. It's not mod engineers. They're all vibe coding. It's senior engineers, senior leaders, people who have, so basically you can, you can narrow it down to 12 to 15 years of experience. They hate vibe coding and they hate AI and they're online going, my 15 years is better than that AI. Okay. You saw, I don't know if you saw Jordan Hubbard's post from Nvidia where he just laid out some really nice advice on how to get the most out of agents as your coding. And this guy posted and he's like, yeah, you know, no, you stick with your

Starting point is 00:01:55 doing your director stuff and leave the programming to programmers, right? When you have 15 years of experience like me, then you're qualified to talk to me, right? Right? So I said something to him, like, I think you need to learn to read a clock. And he's like, and until you have 15 years of experience, and I'm like, well, you got more experience than him. Or I have 45. So should I like go to 60 before I can talk to you? Or should I like cut out 30 years of experience so I can be as dumb as you? Those are my options. And so I don't know. I guess I'll see him in 15 years.

Starting point is 00:02:25 So, okay, I think there's one element that I'm trying to figure out of, while these people have to coexist, right? And most companies are going to have a mix. Even Open AI, by the way, like we talked about this last night at dinner. Guys, open AI has people who don't use AI to code. They have people who don't use codex. They probably are using cursor or something. Okay. But they're not using the agenic loops, right? Yeah, yeah. And yeah, it's, so, you know, we talked to, you know, Andrew Glover there, you know, the director of Deb Prad. And from what he was saying, And they've been planning on going public with this once they have more data about it. Yeah.

Starting point is 00:03:09 And anecdotally, they're sharing that... The performance. The performance difference is like 10x by any way that you measure it. lines of code, commits, business impact, whatever. And it's so stark and pronounced that the people who aren't adopting it, are now 10 times less productive at performance review time. Two people, same title, same job. And all of a sudden, one of them is 10 times it's productive than the other one. What do you do? And the answer is, you panic. You actually go to HR and you go to legal and you're like, what are our options here? Because the time is coming. Okay, here's another hot take.

Starting point is 00:03:34 All right. If you're still using an IDE to develop code by January 1st, you're a bad engineer. There's a hot take for you, right? Now, you still have a, what, five, six weeks to still be an okay engineer while you're using your IDE. But this is the time that you need to drop it and learn how agents code. because it's a skill set. I mean, it's so complicated. We wrote this book about it, me and Gene Kim,

Starting point is 00:04:01 because we were, you know, we were playing with it ourselves last year and we're blogging about it and talking about it. And every blog post was 30 pages. And it's like, what are you going to do with a 30 page blog post? That's long even for me, right? Yeah. And at some point, I was just, man,

Starting point is 00:04:13 like the skills that you've got to learn in order to get the AI to do the things that everyone's mad because it's doing them, right? Because everybody's like, well, I tried it. I spent two hours with it and all it produced was garbage. And the answer is actually you have to spend 200, hours with it. You have to spend 2,000 hours with it. And that's not actually an exaggeration.

Starting point is 00:04:30 Gene just pulled up a study that showed that you actually have to spend a year or 2,000 hours with AI before you trust it. And what does trust mean? Trust in this case specifically means before you, as a user, can predict what it's going to do. And if it's unpredictable, of course you're going to be mad. But as soon as you've worked with it for a full year to where you fully understand its capabilities and its drawbacks, which haven't really fundamentally changed. It's gotten more capable, but the edges are always the same. He hallucinates. It gets lost. It gets amnesia. Dementia. It lies to you, whatever, right? Those skills, we've been building them for years now. Everybody who's been trying to write code with AI, we've been trying. It hasn't really worked,

Starting point is 00:05:10 but it's been working better and better and better and better. And now it's reached the point where it's working a lot better than all of the other options. Yeah. And if you haven't tried it in two months, you're way out of date. The models are much better than two months ago. If you haven't tried it in a year, you're a dinosaur. It's just unbelievable how bad you are. And, and, you know, you may be, look, I have friends who are much better engineers than I am. Okay, I mean, world class, maybe some of the best in the whole world, okay, have built technologies that you've heard of. And they're not using AI yet, except the occasional, I'll ask cursor a chat question like Wikipedia, whatever, okay? Those people are going to be the interns in a year.

Starting point is 00:05:49 You really think so. You have all their experience that they have. So I've had this hyperbole. that has not been really confirmed with any anecdotal evidence at all until today, when I met somebody at your conference that told me about how he had been in this position, 12 years of experience, didn't want anything to do with AI, and he met these two PhD students from somewhere in Europe, I forget where, and they were both just super hardcore vibe coders, you know, with the agents, right? And he was watching them work, and they were super junior, and they kind of didn't know what they were doing, but they just had no fear and all the ambition. And all they did was they just kept hammering on the thing going, okay, well, why did you do it that way?

Starting point is 00:06:27 Explain it to me. Okay, well, let's look at other options. And they would just be kind of the perfect engineer with no context. The perfect, no context engineer, what questions are they going to ask? Have you thought about scaling? Have you thought about security? How is your test coverage, right? I mean, engineers are going to all ask the same questions, right?

Starting point is 00:06:44 And he realized that engineer in a box is not too far off from knowing the right questions to ask an LLM. and that these two students were so productive with it, he was blown away, that he was like, oh, no. Like, that's when the light bulb warnings, I have to learn this. And now he's been doing it ever since, right? But it ain't easy. You're not going to pick up Claude Co. And you're not going to just try it and be like, it's just going to work for it. It might, you might get lucky.

Starting point is 00:07:07 But eventually, if you don't have the right mindset, if you don't have the right attitude going in, even with the right attitude, how often have you sworn at your agents in the last two days with the actual F word or like, right? I'm pretty polite. I say thank you and please. I say thank you and please. I'm going, what did you do that? Right? And it's because, it's because Gene and I realize this after we published the book. You have this helper. They're very human-like. They come in. You have to tell them a lot of stuff and they need a lot of guidance. But over time, they need less guidance. Your prompts get shorter. Things get streamlined. They seem to get it. They're working. Now, if this were a human being, you would draw the conclusion it's because they understand you and they get you

Starting point is 00:07:47 and they're finally part of the freaking team. Do not make that mistake with LLMs. Never make the mistake of anthropomorphizing in LLM, like Larry Ellison, right? The LLM at any moment can stab you in the back, okay? It can just be like, yeah, we took care of that really hard problem. Now I'm going to delete your database, and you're just like, whoa, no, right?

Starting point is 00:08:06 And it's because of that, we call it the hot hand. You sort of like, you're like, it's going, man, I'm feeling good. This thing gets me. I'm going to make it do a production change. And that's how I found out about this. And it's like, I was like, my script can't access prod. And so it chose to do it in the worst imaginable way. What it did was lock out the entire rest of the universe

Starting point is 00:08:25 and couging my live game and everything else and only allowed my script to access prod. And it was changing password. It changed the password. And I was like, why did you change my password? Right? And it's like, oh, I'm so sorry. I definitely shouldn't have done that.

Starting point is 00:08:38 What was it in? Okay, and I'm just, oh, right? This is what will happen to you if you just, you just try to do agentic coding. Okay, bad things will happen. This is what our book is about, really, right? Well, I mean, that's not the best ad because then what? You learn, and then eventually you learn how the speed bumps and the corners and everything.

Starting point is 00:08:57 It's like driving, right? It's like driving. Like, you want to become, like, a NASCAR driver. Like, this is high-performance stuff. You're coding with 12 agents at a time, and you're more ambitious than you've ever been. I was talking to a guy today who's got way more projects going that I've got. I don't know where he gets all the time from, but he's probably doing. 10 or 12, like, major projects at the same time right now.

Starting point is 00:09:20 And he's just doing it all with the gentic coding, you know? So, I mean, like, man, the, the ad here is that you will turn into Batman, but you can't just grab the suit and put it on and be like, I'm Batman. You're just a cosplayer. You're cosplaying at vibe coding. You got to learn how the tool belt works. And that's going to be pain, suffering, and mistakes and learnings. Now, you can get a lot of it by reading this and all of the other vibe coding books.

Starting point is 00:09:45 Read the O'Reilly. Watch the talk. I mean, seriously, like, you should, like, get all of the possible angles at it because it seems to land differently for different people. There will be some analogy where you finally get it. You know, like, oh, I get it. It's like this. And it's like a 3D printer.

Starting point is 00:10:00 And nobody else thought it was like a 3D printer. But somehow that was the magic that made it for you, right? Yeah. I would say one of the biggest surprises from the dinner yesterday was how many people all have the experience where they no longer write single lines of code. Like, they were really just kind of prompting and doing it. quite good by that. Single lines of code? You mean they never write any code at all? They might edit, but like I think when they're writing net new, they always start the prompts.

Starting point is 00:10:24 No editing. No editing. It is very expensive when you're like, that identifier is misspelled and it's a local, you know. You could just add it, but it's better for you to close your IDE and probably uninstall it. No, actually, that's not true. Somebody finally convinced me that IDEs are fantastic, intelligent in particular. Keep it open. It's a cradle build. Yeah. And actually not for the LSP, although you can't use it for that. Actually, that's another good way to use the LOM if you get an MCP server. But no, it's that IntelliJ's auto-indexing is so much faster, and incremental rebuild is so much faster.

Starting point is 00:10:57 This is Radover from last night, yeah. Yeah. So all you do is leave IntelliJ running, but you shouldn't look in it. It's a tool for the AI now, right? Amazing. One other thing that is a big part of the Zlmda hot things you're saying, you say clog code is not it. Claude code ain't it.

Starting point is 00:11:12 Explain yourself. All right. Everyone here loves clog code. Everyone here loves CloudCode. Or AMP, if you use our product, which is just recently leapfrogged Cloud Code again because of Gemini 3. Amp has this cool feature where it goes to another model. Just to pre-warm you, I also want to talk about just Google in general and how this Gemini revolution has kind of changed Google's image. But let's talk about CloudCode.

Starting point is 00:11:35 Sure. CloudCode's been around since March. CloudCode has been proven to work. But yet probably 80% of the world's... 90% of the world's programmers are not using it or anything like it. You get certain companies where it's really taken off, you know, but most aren't. The world is stuck on cursor. The world is stuck in 2024. Last year, we were trying to get people to write with chat, right?

Starting point is 00:12:00 And we were like, and they were like, no, completions. We were like, oh, God, no, but it can generate the code. You just got to paste it in. You just got to do all this stuff. And they were like, that sounds kind of hard. And we're like, but it's faster. And they wouldn't do it. And then nine months later, it finally percolated in.

Starting point is 00:12:16 And now they're all like, I like cursor. And it's like, that's so last year, dude, right? Like, wake up. And yet they haven't adopted it. And so you have to, at this point, look at it and say, why haven't they adopted it? Let's go look at the reasons. And the answer is, it's too hard. It's too hard.

Starting point is 00:12:30 You have to be able to read, man, most engineers, honestly, like, to them, five paragraphs is an essay. Okay? And with Claude Code, you've got to read waterfalls of not just information, but also. code and diffs, right? Because if you're going to put your IDE away, you actually do have to look at the diffs. Now, I'm going to tell you that once you get some expertise at this, you can actually tell from the shape of the diffs and the color of the diffs and the length of the diffs. The vibe. You can tell whether it needs a code review, whether they're doing the wrong thing, whether they seem to be rerouting suspiciously too much code for this problem. Right? The diffs alone,

Starting point is 00:13:02 just the shape of the diffs can tell you a lot about what's going on without actually reading the code, but you should pay attention to them. Otherwise, you'll have problems that will only crop up later, right? But yeah, I mean, like, uh, uh, uh, uh, uh, put the IDE away. Okay, clog code, and then get clog code out and try to start using it, all right? And you're going to find that it's,

Starting point is 00:13:22 look, I've been using clog code honestly 10 to 12 hours a day, literally, for months and months and months and months, and I still curse it out all the time. I just lose my mind. I'm like, how could you have done that when you just set, right? And it's like, it's actually been shown,

Starting point is 00:13:37 it's starting to be shown that sometimes when you put a little pressure on them, they perform better. You can break through law jams that way. But anyway, look, you're going to run into problems. But the thing is, next year the tools will be better. Okay?

Starting point is 00:13:48 If cloud code's not it, what is it? Well, we got to get back to something like an IDE, right? I mean, that's just going to be, it's got to be natural for people. You've got to be able to look at it and see what's going on, not have to read. It's got to have visual indicators, right? And yet it's not going to be an IDE because an IDE is very much focused on helping you write code, and that's not what you do anymore, right? So what it's going to be is it's going to be your agent orchestration dashboard.

Starting point is 00:14:11 You're going to walk in in the morning and be like, yo, So, how our engines do, right? It's like, oh, that one's still running. That one's running a tool. That one needs my input. Okay, right? You just go through the list. And so I'm building one.

Starting point is 00:14:24 You can go look. I'm supposed to be a private repo, but it's public, so I've got forks and shit. Happens. But whatever, you can play with it. It's called VC vibe coder. It's my B2 of the vibe coder system. And what it does is it creates a set of scanned workflows that run the agents for you. Yeah.

Starting point is 00:14:38 I don't know if you saw anti-gravity from Google the other day, which you all two days ago. It's so fun how much stuff people are inventing that are all They're ageo marries Yeah, so look, I called this, I don't know, I called it in March With Revenge of the Junior Developer, I did that chart and everything And like, Dario quotes it and all this custom advisory boards and everything, right? Really?

Starting point is 00:14:58 Yeah, yeah, no, it was really pretty impactful. And I called that what's going to happen is the agents, even back in March, I knew they were too hard. I was like, what's going to happen is they're, you can run them programmatically and 90% of the crap that you do with them could be handled by a model. often a cheaper model, right? If it's just like, if it's asking you, which of these two things should I do next

Starting point is 00:15:17 that are equally important? Like, just have haiku say either one, right? So, like, I called the orchestrators are coming and it's taken close until, like, the end of the year to get there, which is roughly where I predicted them coming. Replit, Agent 3, there's a bunch, there's the conductor, there's a D-Mad came out open source. They're all different, you know, takes on it, right?

Starting point is 00:15:37 And, but there would be more coming. I guess Google's as well, right? I like this analogy that they have. It's still pretty new. So who knows what the eventual vision is, is that you just get notifications from your agents as they're working. Exactly, yeah. So in mine, in VC, there's an activity feed.

Starting point is 00:15:53 That was one of the first features I added, which is like, I wanted to go work, and I just want to get notifications periodically of interesting stuff. Interesting. I wonder if they'll have, like, social networks of agents. Well, so the agent-in-in-eat-e-e-all-ing-e-all-all-all-all-so I just had three-hour The coffee with Jeffrey Emanuel, who he did the MCP agent mail. He's one of the smartest people I've ever met in my life.

Starting point is 00:16:15 He's the one that wrote the article that crashed the Stark Market about Nvidia. That Jeffrey Emanuel, an incredibly well-written article that said, this is why it's a bubble, and the whole market went, and Carpathy started following. It's back up. He wrote what you just said. It is back up. But he wrote Agent Mail, which is he was just tired of having to copy stuff between his agents. Like, you tell me what to tell this agent.

Starting point is 00:16:39 And so he made a little, like, I don't know, HV server that's like an inbox for them, a messaging. And they talk to each other now. And now he goes, coordinate amongst yourselves to parallelize this task, this epic that I just put together or whatever, and they'll do it. Some people are coming at it top down

Starting point is 00:16:53 and trying to build orchestrators that you do it all for you. But interestingly, with beads, right, which is the issue tracker session thing that I made, plus his... Purely vibe-coded, by the real. Purely vibe-coded, yes. So, I mean, like, I get PRs every day

Starting point is 00:17:05 for horrible problems that I introduced, but nobody seems to mind because we've got stable versions now. So Veed's is like living proof that you never actually have to look at the code as long as you and other people are asking the right questions and having the AI look at the code. I get PRs from people all the time where it's obvious that the AI did all of the analysis and all of the coding and I look at it and sometimes I'll just be like, so my AI, what do you think of their AI is PR, right? And you know, it's all summarization.

Starting point is 00:17:30 I mean, isn't that bad? Don't you want... It's bad if your code. Look, it's all about the outcome. Bades is working and it's got tens of thousands of very happy people use it. So obviously it's not bad. I mean, one of the supergives out there. If you do this to your company's production website and bring it down, then yeah, it's bad.

Starting point is 00:17:47 But still, Bates is kind of a database, you know, and database is one of the harder things to make. You know, Beds is really weird. The architecture is really weird. And the only reason it works is because it wouldn't have worked in the old days. It would have been just too hard to manage and not programmatically. But what you do is you tell the AI, go fix it all up. And whenever it's corrupted or there's a merge conflict, or just fix it. And it's funny because Jeffrey Emanuel, who did the mail, basically did the same thing.

Starting point is 00:18:11 He has all his agents run in the same directory. And they do file reservations. They're like, I need that file. Man, I used to do that Accenture in the 90s, right? I'd like run over to a dude's cubicle and be like, I need that file. Their revision control was so bad. So, like, he's got a file reservation system going. But what happened was as soon as he put it in place, his agents just started working.

Starting point is 00:18:32 And now he's got this little village of agents, right? And that's where we're headed. So the orchestrators are going to be about not keeping the agent on the rails, but keeping all of your agents on the rails and communicating with each other. And then you hit the wall. Boom. Does anybody know what the wall is once you get past all this? Merge.

Starting point is 00:18:51 Merging is the wall that everyone is hitting right now. I think the company that's best poised to solve it is Grappite. I was going to go talk to him about it. they'd be happy to talk to you, yeah. Yeah, I think everybody needs to solve it. And if you're at an enterprise, like what we hear, because Gene Kim and I talk, we talk to companies all that. I'm a SaaS seller in a source graph,

Starting point is 00:19:13 so we get to hear the inside story from all these big companies, right? And they're saying, yeah, as soon as you get to the point where, like, every developer is 10 times as productive, merging their code becomes this incredibly complicated problem because you and I work at the same time for two or three hours. We make, you know, 30,000 line change each. Mine makes it in first, and it gets merged, and then you come along. And I have literally changed our logging system and our, like, you know, our architecture here and APIs that you are using.

Starting point is 00:19:44 Yeah. And so it's not going to be a simple, it's not as simple, let's fix the merge conflicts. It's like you're going to have to re-envision and reimagine and reimplement your change on my change. Or rip yours out. Or rip mine out and make me do it. But ultimately, ours are just the AI's different. doing it, right? But the important thing is that they have to be serialized. It is a queue, and that when they go in there, they have to actually, like, basically redo what they were

Starting point is 00:20:09 doing on top of the new thing. Nobody has solved this, and it is a huge obstacle right now. You know what one company did? Sorry, last thing. One company said, here's our solution. One engineer per repo. Not making that up. It's a solution. It's a solution for now. The classic solution for this is stack diffs, right? Merge Q's stack diffs. I don't know about So I guess I'm dumb. It's like a Facebook concept that they're trying to bring into the wider roles. GitHub is working and adding it. I just talked to Jared Palmer there.

Starting point is 00:20:37 Basically, I'm hearing no solution yet, but you should be aware of it and design around it. Yeah, I mean, there's the old-fashioned way of just hammering through it really hard. Well, also, you know, you could just talk to the other guy and say, like, hey, I'm doing this, you know, pretty deep architectural change. Let me go first. And let's agree on the overall pattern first. So, yeah. I mean, I've run into this situation a few times where I've actually tried to give this agent the heads up that this one's making you change that affects this one.

Starting point is 00:21:00 With the mail thing that Jeffrey did, I think once I get it wired up, because he doesn't use work trees and I'm going to, this. But once they can actually talk to each other, I think it's going to be as simple as, just keep in mind that that agent's working on something that affects you. You might want to go talk to them about it. Yeah, and agree on the overall, like, fundamental if, and they're quite good at it. They just, yeah. It's because they have no ego.

Starting point is 00:21:22 They're not like, oh, it's got to be me, right? So just whoever's first gets to be the leader. Great. What do you and him disagree on? Me and who? Jeffrey? Emmanuel, the guy that I just met. Well, we, so we foundationally, fundamentally disagree that having 12 agents to work in a single repo clone is a good idea.

Starting point is 00:21:41 So you're on the pro side. I'm on the pro lots of, like, either get work trees with lots of branches or separate repo clones. I would imagine. He's them sandbox. He's in favor. He's got them all in the same. They're all, they're literally, they're using the same get, the same build. So one of them will be like doing a build, like need to run a test.

Starting point is 00:21:58 That's so much churn. Yeah, but he has a file reservation system. So the funny thing is, okay, I was like, this is insanity. And he's talking me into at least acknowledging that it probably works pretty well if you're a solo dev and you're using no more than a dozen or 20 agents because it is actually working for him. And he uses the same principle that Beads does, which is it wouldn't have worked in the old days. It doesn't make any sense to a real engineer. And yet you tell the AI, if anything gets messed up, just fix it.

Starting point is 00:22:24 And they will. And so that's, right? That's why his thing works. Because every once in a while, the file reservation gets screwed up. And they're like, hey, we need to resolve this and they figure it out. Interesting. Yeah. Some people have proposed that the theme of this conference next year is on multi-agents.

Starting point is 00:22:39 Oh, yeah. I mean, yeah, of course. Yeah, yeah. I mean, AI will be about multi-agent. Look, we're in this phase still where we're cutting down corn with siths with our hands. That's what a real programmer does these days. We're moving next year. It's very clear.

Starting point is 00:22:52 We're moving to, you know, these machines that turn. You know, these giant, just like those ones that you see on the farms today, factory farms, we're going to be factory farming coat. Okay. And that absolutely, like, a lot of people are just so dead set against that philosophically, morally, ethically, whatever. They're just like... They're so used to subsistence agriculture that we're not used to like the big John Deere.

Starting point is 00:23:17 But we are actually moving into the John Deere era of coding. That's amazing. Yeah, but the funny thing is actually. And I just thought of it, too. We'll have to reuse it. Yeah. But it's been growing on me. It's the whole, it's this idea that Claude Code Code and AMP and Codex, you know, Klein, we love them all.

Starting point is 00:23:35 Equally, they're all equally bad. I said in my talk today, they're like, they're like a power saw or a power drill. A skilled craftsman can do a lot of good with them, and then you can also cut your foot off with them. The same thing's true of Claude Code. But imagine a big machine, a big farming machine that knows how to run Claude Code and scrub it, right? Almost like, it's like, okay, you plan, you implement, you. review, you test, right? You split it all up. And now you got yourself factory farming, right? It works. People are building it. It's going to happen. And what it's going to do is it's already

Starting point is 00:24:06 started to unlock programming for non-programmers. And this is completely turning companies upside down. They're starting to realize that maybe the ideal team size is like two or three. And I mean like, right, the whole way that companies are run, the whole governance structure is going to change because now coding is no longer the bottleneck. The business needs to get immediately involved. The feed back loops get faster and it's really exciting times, but it's too much for a lot of people and they just, they're like checking out or they're revolting online. And I predict that as this capabilities improve and as we get closer and closer to the factory farming of code, we will see a massive backlash from the Luddites. You are the one of the few people I can ask this as a, I know a lot of people

Starting point is 00:24:47 on our audience are critical of going the full hog with this. Yes. So a lot of like, they're like, Fine for front-end, find for application code, but don't touch my cloud infra. Don't touch my backend, my distributed microservices. Definitely don't touch anything, production. Only touch code. Only use these things when Git is your backstop, for starters. So keep prod out. It's going to be real tempting to, right, but don't.

Starting point is 00:25:10 If you have Git is your backstop, why should you be worried? True, except, I guess, people have the perception that it is less good at back-end code. Oh, this is the problem where everybody's added math. Yeah. Okay, so how good was Chad GPT 3.5 at Systems Code? Pretty bad. How long ago was that? Okay.

Starting point is 00:25:30 Two years ago. People think the, honestly, I believe that the misunderstanding here is rooted in a fundamental belief that the models are done getting smarter. Right. And the funny thing is, they could be done getting smarter. They're not, but they could be, and we would still be over the hump where we've discovered electricity and now we need to harness it. Yeah. We will still get to factory farming code with today's models capability. and we'll get there fast. We'll get there by summer. But the models are getting smarter so fast.

Starting point is 00:25:57 You know, it's really, there's this interesting tension of, you know, like, you're building tools for capabilities that the models will eventually have built into their brains. Yeah. And so you won't need that capability in the tool anymore. And so there's this constant arms race and decay of your tool filling gaps for the model until the model's good enough to fill it itself and then your tool moves on. Yeah. All road is becoming, all code and all tools are becoming throwaway. Yeah. Which is great because they're easier to build too. Yeah, by the way, yes. So remember Joel Spolsky, one of our generation, our time, one of the greatest writers and thinkers, he gave the best tech talk I've ever seen, and I want to get him to come and revive it. He gave it at Amazon 20 years ago. It's still relevant today. Great. So Joel Spolsky, a long time ago, wrote something that was timeless until today. So it was 20 years timeless, which was never rewrite your code.

Starting point is 00:26:44 And now we've discovered that it is for a larger and larger and larger class of bodies of code, it is better to just start over and rewrite it from scratch than it is to try to fix it. The LLM will do a better job. I first noticed this when I was trying to port all of my unit tests from one architecture to another. And eventually, it's just an iteration because they're trying to fix. There's a lot to keep in hot. But instead, if you say, throw all the tests out and make them again, it just goes and you're done. Right?

Starting point is 00:27:10 And so it's like, hmm, well, what about this library? to refactor. And so it's creeping up. But we're moving into a world where the fastest thing to do is just build new code that does a better job of what the old code was trying to do. Yeah. I mean, it's like we're unlearning everything. I feel like an upside down land. But this is, but it's like we've entered quantum mechanics. But you have to, you have to embrace this new world. I love the energy and the credibility that you bring because a young kid could say what you're saying and not be as believable. But you're coming from the perspective of you've been a huge system program, you've been a game program, and you've been everything.

Starting point is 00:27:43 Yeah, I've done assembly language for five years, you know? Yeah. Operating systems in assembly language. And it was 80, 80, 86, not even 80X 86. We had 8-bit registers. I've done it all. And, you know, the game programming teaches you everything. And then, of course, I've done platforms, Google, and ads, and this and that.

Starting point is 00:28:02 And, you know, the agenic loop and the game programming loops share a lot in common. They do. Resource sharing, operating system loops. I feel like I'm building the same systems over and over again now. Yeah. It's, it's only, we're cursed to reinvent the same designs in every new domain. It's a privilege, too, you know. One thing I wanted to get you a comment on is Google.

Starting point is 00:28:22 Oh, Google. One of my favorite memories, which is like just before you retired, was talking about how Google still doesn't get it. Google Cloud in particular, how they shut down. The deprecation policy. The deprecation policy. I was so mad about that. You got to get me pretty mad to write a blog.

Starting point is 00:28:36 They seen, have they turned it around? No. I talked to some people there, and a lot of them were like, yeah, that's not a thing for Google. And it's funny because, you know, Amazon, not on the platform, not on the deprecation stuff, not on the important stuff. Google has turned it around on execution. They finally did the thing that they should have done, you know, 15 years ago, which is hold people accountable. And it's not just engineers do whatever they want all the time, which was what it was for 20 years.

Starting point is 00:28:59 It actually worked pretty well because they had a monopoly on ads and they could afford to subsidize Google engineers doing whatever they wanted for a... But, you know, ultimately they had to do the right thing and grow up and mature as an organization. and it was painful and they lost some Google culture and it's not as fun anymore, but they now execute well and they did the right thing for the company. And now with Gemini, you can see now they've been shifting their focus gradually towards more AIAI and now it's starting to pay off for them.

Starting point is 00:29:23 And maybe they're going to be the big, big winners. Do you have observations of a similar kind with all the other labs? I'm just kind of curious and your takes on one of my favorite charts is that old chart where you had Microsoft all pointing guns at each other, Facebook, everyone's a ring. person to ask me this, I remember that chart, that was funny.

Starting point is 00:29:42 Yeah, I feel like someone could do that for opening eye. They could, they could. You know, it's an interesting question. All three of those companies, Google, Anthropic and Open AI are an unbelievably chaotic internally right now. Yeah. Chaos. Okay. Anthropic hides it really well.

Starting point is 00:29:58 They seem like they've got their act together. So what that means is their product managers formed a wall around that chaos and Bravo, Anthropic product managers. But it is, and it's not because Anthropic screws. it's because it's an inevitable function of growing that fast. They're hiring like 100 plus people for Claude Code in the next, I don't know, month. I mean, like, they're going wild. And that's just Claude Code. You're not going to, I mean, I was at Google and Amazon when they were in the big fast phases and you're just going to have chaos. You're going to have churn. Nobody knows who

Starting point is 00:30:26 to talk to what and everything's crazy. Eventually it starts smooth out, settle out, and they'll get there, right? Open AI is chaotic more like in a, well, they had a lot of exits, right? You know, I don't know if there's chaotic to say GitHub who lost most of their senior leadership and was just complete turmoil for years but they're pretty chaotic at OpenAI right? And then Google you know

Starting point is 00:30:47 we were just talking to somebody today that was saying it was still too hard to get consensus across groups with the Jules team. They can't get it rolled out internally because Google is so siloed. It's a billion monoliths, right? Little little apps that don't talk to each other

Starting point is 00:31:02 that it's hard to roll anything out across Google. So all three of them have execution problems right now. I think Anthropics probably executing a little bit better than the other two, but it's real close race. And yeah, it'll be interesting to see. And see if Oracle or Facebook or any of the others can catch up, right? Meta. Facebook will be the most interesting thing. I mean, they'll have to do something huge next year. Next year could be the year of open source models. Yeah? Well, so, look, as soon as open source models get to the point where there's as good as CloudSsonnet 3-7 was, then you turn on Klein or something and you've got something that as good as Cloud Code was in March,

Starting point is 00:31:35 which wasn't as good as today, and it's not good, but it's good enough, and you're running it for free, free, free, free, on your local M4 or whatever, right? So, yeah, and from what I've heard, they're seven months behind, and that gap is gradually narrowing the frontier models, which means OSS models will be as good as Gemini 3 next summer. Right. So, yeah, next year could very much be the year. That means the tools are going to have to get much, much better at decomposing the task

Starting point is 00:32:04 and assigning them to the right model, the right size of model for cost optimizations. I'll represent the critical side, which is that the reason they're converging is because they're saturating, right? You can only ever hit 100, and the closer you get to 100, proportionally, it'll just get harder and harder, right?

Starting point is 00:32:20 So obviously the rate of change when you're lower down is higher as compared to when you're already saturating. But that's a minor technical point. Well, no, I mean, it's not minor at all. It's actually a foundational question, which is, is the line of AI, intelligence going to go straight or is it going exponentially or is it actually starting

Starting point is 00:32:38 to be... Ascentotic, yeah. Yeah. And, you know, from what we've heard from people who are very, very close to the research, we know that AI has been getting, what is it, four times smarter every 18 months for the last, I don't know, 30 years because of Moore's Law. And they think that there's enough data left,

Starting point is 00:32:54 training data, for two more cycles of that before they don't know what happens. Yeah. Maybe it goes up more or maybe it goes down. We don't know. Human history ends. But two more cycles means they're going to be 16 times smarter in three years, right? So, well, I don't even know what that means. Well, I've spent a long time trying to figure out what it means, but what it means is they're

Starting point is 00:33:13 going to be really, really, really, really smart, and it's going to change the world, probably in a lot of good ways and a lot of bad ways. And, yeah. I don't know if you have this version in this conversation. People ask me if their kids should learn a code. Kids should learn to vibe code. You, like, you have the escape hatch of, you can read the code if you want to. You just don't need to, most of the time.

Starting point is 00:33:34 But you can, and it's a good guard. Right, but I don't because you don't have to. Well, I think my take is whatever it is, you'll be better off if you do also know how to code because you can prompt better. Right? Because you can tell, you can communicate or precise terms. Look, when I say you know how to code, not the syntax and stuff, but you have to know, like, in a language neutral way, what the capabilities of languages are. functions and classes and objects and I don't know monads whatever it is the whole super set you should be aware of them and then from there up so you you've you've cut off all the syntax you don't care how to write it anymore but you care how it works

Starting point is 00:34:14 so you've sort of reached the level of how a product manager thinks about things architecturally right and you need to be that product manager and now you're starting to move your concerns up up and you need to know all the engineering stuff and like Jeffrey Emmanuel like I was talking about he's a mathematician self-taught engineer he doesn't but he he's learned all of the right concepts, you know, you know, Cloudflare does this and Apache Cassandra does that. Yeah, that is still technical, yeah. That doesn't go away. You still need to learn all that, right? And so just because you don't have to write code anymore doesn't mean you have to, you still have to learn a massive amount of stuff to be an effective engineer in the new world because that's the level that you're interacting with the mat. Amazing. So this has been a great overview. I don't know if you have any other sort of rants in you that you want to sort of get

Starting point is 00:34:57 out there. I'll leave you the floor. I feel like the gossip rate has gone up, like not gossip, but the rate of exciting announcements by engineers who have discovered new things about how to be more productive with agents. Like, for example, I just found out today, not this, I found out today about it's called code MCP or something like that,

Starting point is 00:35:15 where you, uh, instead of calling... Pretty popular projects. The agents can't call MCP very effectively because they don't have any training on tool calls, but they have plenty of training on writing code. So you tell them, don't call the tool, write code to call the tool, and they do way better with it, right? So it's like, it's all these little learnings that we're finding, right, are just...

Starting point is 00:35:31 It was crazy to Anthropic, the creators of MCP found this. Did they? Yeah. Well, Cloudflare found it first, but then Anthropic was like, yeah, yeah, you guys are right. Yeah, wow. That's really neat. So I think that's why I love focusing on the AI engineer, because my argument is the AI engineer can uniquely take advantage of LLMs way better than everyone else. That's true.

Starting point is 00:35:51 So you go so much more problem. You could almost define an AI engineer as somebody who's mastered LLMs. Yeah, yeah, yeah. Not from training, but from using. Yeah, yeah, yeah. I think it's not. don't these disruptor, like, strategies where, like, it's low status. It's high status to be a researcher.

Starting point is 00:36:07 It's high status to train models. You don't get any respect if you're a GPT rapper. But, like, people are starting to be more productive and, like, actually develop sincere expertise in the same way that, I think, like, F1 car drivers don't know how to build an F1 car, but they know they'll tell you everything about driving it to the... And they may know, in a sense, they know more about operating it than the people who build it. And so they have to have that conversation, right?

Starting point is 00:36:31 Yeah. Although if you watch the, I think, the F1 movie, you get a little set of the other. Oh, and they make all the money. Is that what you said? That's good point. It's flip-loft. Lovely. Well, thanks so much for coming on. A huge reminder of your work. Your energy is very infectious, and I hope you keep doing Stevie's Tech Talks. I'll start them up again, man. I mean, this energy is because of the AI and it's because of vibe coding. It's addictive and fun. Tech is fun again. It's going to go boring for a little bit. I know. I know. For a while, it was like, well, Sourcegraph, like, index your, your, your codebase, like, really, really well, you know? And again, it's, like, so, so fast.

Starting point is 00:37:06 And I'm like, well, that's cool, but, you know, what's cooler? It's not good. It's cool. It's been fun.

Latent Space: The AI Engineer Podcast - Steve Yegge's Vibe Coding Manifesto: Why Claude Code Isn't It & What Comes After the IDE

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.