Geoffrey Huntley argues that while software development as a profession is effectively dead, software engineering is more alive—and critical—than ever before. In this episode, the creator of the viral "Ralph" agent joins us to explain how simple bash loops and deterministic context allocation are fundamentally changing the unit economics of code. We dive deep into the mechanics of managing "context rot," avoiding "compaction," and why building your own "Gas Town" of autonomous agents is the only way to survive the coming rift.
Show Notes
Geoffrey’s Website & Blog: ghuntley.com
Build Your Own Coding Agent Workshop: ghuntley.com/agent
Ralph Wiggum as a Software Engineer: ghuntley.com/ralph
Steve Yegge’s "Welcome to Gas Town": Read on Medium
The "Cursed" Programming Language: github.com/ghuntley/cursed
Transcript
(Disclaimer: may contain unintentionally confusing, inaccurate and/or amusing transcription errors)
[00:00:00] Andrew Zigler: This is so exciting. I'm actually kind of excited that we also have those lost tapes because those, that was some really good shit that you were putting out there a second ago
[00:00:07] Ben Lloyd Pearson: Those are our secrets. Now we get to
[00:00:09] Andrew Zigler: And now those are like, wow. Those are now, now only, I only I heard that, uh, that was a closed room. That was great.
[00:00:15] Andrew Zigler: But I, so this is just like overwhelmingly cool Geoffrey like to have you here, you know, just to give you some context, we've been talking about you a lot on Dev Interrupted for the past year. I would attribute really what I know about starting to work with these tools and even using Cursor to you. All of what I've learned in the last year started with reading your articles as like the nucleus of that thought.
[00:00:41] Andrew Zigler: And so to be here and be able to kind of chat about your ideas and how they've evolved is really exciting for me. Because I watched you bring the, the now very well known Ralph into the world. Um, and that's what we want to talk about today. And what it means for developers that are maybe using their [00:01:00] own Ralphs.
[00:01:00] Andrew Zigler: Because I think there's a, a huge revolution happening right now at the top of the year with all the engineers coming back from the break with new ideas. Opus 4.5 under their belt, and just a lot of, insight from folks like you, um, out in the field that are experimenting. So we're really stoked to have you here.
[00:01:18] Andrew Zigler: But I, I, I wanna just also frame the bash loop itself and kind of take a step back because I wanna start there and just ask you when you brought Ralph into the world and you named it Ralph, why, why give it that name?
[00:01:32] Geoffrey Huntley: Yeah, so it's February last year, almost a year ago. And I was doing spec based, uh, development. I made a mine at GitHub next published actually, uh, some research there about two years prior to that. And I remembered that, um, like spec based workflows or l LMS generating apps. This is like the GPT-3 three days.
[00:01:59] Geoffrey Huntley: So I [00:02:00] started applying it and then I started realizing that like, I could probably, this could be quite mechanical, like if I take specs and what, how I'm, I'm doing things and then I, I got to the point where I was real like, wow, there's some pure first principles thinking here of like, I could deterministically allocate the array. And the first thing I was like, okay, what can I do with this? And this was the Z 80 blog post. And I started the path of essentially, uh, using LLMs to essentially clone product features and of existing companies.
[00:02:43] Andrew Zigler: Yes.
[00:02:46] Geoffrey Huntley: Like taking source code that was like BSL license, source code and like clean rooming it into specs.
[00:02:56] Geoffrey Huntley: Very similar to how the Intel and AMD CPU came into, [00:03:00] and I reimplemented like HashiCorp Nomad, I cloned the company Tailscale. I did all these random
[00:03:09] Andrew Zigler: remember Rovo we read Rovo. You,
[00:03:13] Geoffrey Huntley: And all
[00:03:13] Andrew Zigler: did several of them.
[00:03:15] Geoffrey Huntley: Yeah, this is February
[00:03:15] Andrew Zigler: It was, it was really interesting to watch how you used the LLMs to do this, like a binary analysis and then you would pull out this information or rebuild what they had put in it. And it really kind of took away, uh, it like, it, it drained any moat that you could find.
[00:03:29] Andrew Zigler: Like you could go to any moat and drain it with that and be like, I can build this in a day or with this loop and figure out how this works. Nothing can stop me if it's on my computer.
[00:03:38] Geoffrey Huntley: Pretty much so for developers, you go speak with other founders and how they're, they're feeling about this. Because that, it's a, it's pretty gloom out there for founders. Like, we can focus on the developers and developers side of things, but go speak you founders. Now, to answer your question, when I [00:04:00] discovered this in February, it literally made me wanna, Ralph.
[00:04:04] Geoffrey Huntley: Ralph is a, a term for vomiting. And it wasn't named, it wasn't named, like, it wasn't a, a technique was not named, but it made me want to vomit 'cause I could actually see where we're going. I could see where we were going. Like I was building software in my sleep. And then I started publishing a lot more. Um, I focused heavily in on like, dear juniors, you're screwed if you don't pay attention. Got that distributed to prime gen. So the Gen Z popped off in that area and that was like real guidance explaining the cycles of our industry and, uh, really. focused on some of, you're not gonna make it. You need to be curious.
[00:04:46] Geoffrey Huntley: You need to be curious. There was reasons behind the order of my publishing because I was spooked. I was feeling like really sick in the stomach, knowing what I essentially discovered, uh, that if you drove it in the right [00:05:00] way, allocated the context window, deterministically in a loop, and it was so all so simple that I could see where it could go. But in the, in the future belongs to everything talk I do. And I say like, I walk around, I see dead people like the sixth sense. Like it's not like they're literally dead, just, they just don't know that, that they don't have a job yet. And I'm like, wake up, wake up. Please wake up. Please be curious. Like play with the LLMs.
[00:05:28] Geoffrey Huntley: Like, they're good enough. Now Play, please play. Please play. And then I went across to San Fran, um, went to a meetup, uh, read into Dex Horthy. And we had a very similar moment, so that I had that moment in February to my, with myself. And then in July I just rocked up to a meetup in San Fran. Two hours late, presented like, uh, presented last, and the, we just kept going all night and it was like 15 [00:06:00] people in there and we spent an hour afterwards going, oh my God.
[00:06:04] Geoffrey Huntley: They could see it. They could see it. And because it's San Fran, I decided, Hey, look, this, the word's gonna get out about this thing 'cause I just essentially showed the technique for the first time and I decided, well, it needs a dame. It's kind of dumb, it's kind of lovable and it never gives up. And it made me wanna, Ralph, so I called it Ralph, as in Ralph Wiggum.
[00:06:35] Geoffrey Huntley: And thus, if forever is now stuck in meme folk lore, Um, but it also has an alternative, meaning it also has an alternative meaning, which comes into the future of software engineering. What Dex and I have calculated is the software development as a profession [00:07:00] is dead, but software engineering is very much alive.
[00:07:05] Geoffrey Huntley: Okay? If your, if your value to a business is that you type things on the keyboard would you believe oppa, sorry, sonnet 4.5 on a loop with, with a bash loop, Ralph, the, it costs $10.42 US an hour. Now, what's the minimum wage law? What's the minimum wage laws in your state? I'm pretty sure that's why the discussion in San Fran was going on so long was we realized that a fast food worker will get paid more than a software developer. Now, I don't mean to be inflammatory in saying that, but that the unit economics changed. It changed six months ago. There's been a batch of Y Combinator companies Ralphing. You folks were early. You are reading, you're growing like there, there now exists. People who are year on, a [00:08:00] year in or six months or a year in are even applying it.
[00:08:03] Geoffrey Huntley: And that, that the, the rift on this is becoming huge. Huge. So, so unfavorably huge. It, it, it is unbelievable the difference. Someone who's been curious in the last year and invested in themselves has as an advantage because all the frontier labs were giving away inferencing for almost free. Now we've got these quotas and caps, so we were able to burn tokens into absolutely crazy stuff with this space dust. Now software, software engineering, however. Is alive, software engineering is alive. If you, if you are just doing development or typing the keyboard, yeah, it's predictable. What's about, what's about to transpire. Autonomous worker can do your job. A PM can drive autonomous worker to do your job. Now, software engineering is a different discipline.
[00:08:56] Geoffrey Huntley: This is the, the the like actually [00:09:00] engineering. If the idea of autonomous loop within your code base makes you want to Ralph, listen to that feeling. Autonomous loops, just, it's like the, it's like the shipping container. Like you look at the cost of freight per ton before the shipping container was invented and it was sky high, and then it was like sensed to the ton after the, the shipping container.
[00:09:29] Geoffrey Huntley: The Ralph loop is that same type of equivalence of the change in the unit economics. It's not going back. So if you are, you are, if you are the software Dev or, or the, uh, sailor, like carrying cargo on and off the ship manually, please understand the dock, like the, the, the container, the box just got invented.
[00:09:50] Geoffrey Huntley: Now engineering is different to software development and engineering is now the game because let's say running an autonomous [00:10:00] loop, you're like, oh, what happens when it drops a database sweep? You're an engineer, right? What are you gonna do to stop it from dropping the database? Don't provision, uh, write secrets.
[00:10:10] Geoffrey Huntley: Maybe you introduce some tests and you enable, uh, change, data capture. And then you, you expect the audit log, like engineer your way out of the failure scenarios. And the more you do that, the more you capture the back pressure, which allows, uh, more autonomy. For the new unit economics, that's the game now.
[00:10:33] Geoffrey Huntley: Um, like pre-commit hooks, property based tests, snapshot tests and eventually you'll come to where I am, um, where I don't even use ci. And my agents run full with pseudo on a machine, uh, bare metal box running NixOS. I don't review the code. The agents autonomously [00:11:00] push to master. No branches. And the deployment happens in under 30 seconds. Then if something goes wrong, I've got feedback loops which feed back into the active session and then it just self repairs. So we're starting to get into, when you start to get into some, what I'm researching, I call this loom. and Steve Yegge calls this gas town. We're exploring what it means for to be self evolutionary software.
[00:11:32] Geoffrey Huntley: Now you don't have to yolo, commit into production. There's no reason why agents can't be in control of feature flag experiments and add in the branch, and then they, they connect us to some other data source, and then they make the cut over and they, they roll it all out again. So this is, this is engineering now, folks. It's all an engineering game now.
[00:11:57] Ben Lloyd Pearson: You know, you're, you're, you're talking early on [00:12:00] about the importance of that curiosity. You know, and I, and I agree with you and I, and I feel like it's, right now feels like the best time in my career to be curious. Like, there's so much opportunity when you, when you know where to look.
[00:12:13] Ben Lloyd Pearson: And I feel like a lot of people. Their perception of this technology is a little bit tarnished by what may have been early exper experiences with it. Like, you know, we, we've all been using these tools since the, the days of GPT-3.5 or whatever it was, and, and, and I know that, that the things that it failed on back then were, were obvious.
[00:12:33] Ben Lloyd Pearson: Often. They often I felt like it, it really did not seem like a very effective tool. But you know, here we are. We've, we've had multiple generations of models come out. The models have gotten substantially better in, uh, in the last year in particular, I think. And, and I feel like there's the, because the reputation of AI has been tarnished so much by the early phases of it, that people are struggling a little bit to see like what the next version [00:13:00] of this all looks like.
[00:13:02] Geoffrey Huntley: Yeah, I've got
[00:13:04] Ben Lloyd Pearson: I, I'm, I'm, I'm actually kind of curious if, if you felt like, if you felt something similar, even, even with the, like these autonomous loops that you're building, that you're Ralphing. Are you seeing like the, the changes to the models even like resonate into like improvements for that method of doing things as well?
[00:13:20] Geoffrey Huntley: Oh, absolutely. So the start of when I really started publishing, like from my oh fuck moment, was like Boxing Day, I guess two years ago now, or 12 months ago, really. That was like Sonnet 3.5, and I was able to generate like a Haskell audio library. It was so stupid. It, it wasn't like regurgitating Stack Overflow.
[00:13:43] Geoffrey Huntley: That was my mindset and my frame. Oh, just regurgitate stack overflow. I tried this at GPT-3. It's not anything, and, and I, I came back 'cause I run in, in a loop like by hand in, in Cursor. I wasn't quite yet at the very like, [00:14:00] novel idea of just running in a bash loop. It was just very mechanical. Like I was just doing the motions of specs and everything else, and that just shook me. And then when I got to February, this was at GPT-3. Uh, so this, that, that shook me. My previous frame of reference was like GPT-3 ish. So that was like, it took a call with a, a founder to say, Geoff, you need to get on the tools. Like you need to start writing. Do you see what I see you, I'm gonna hang up now, and you go play and you call me back.
[00:14:39] Geoffrey Huntley: 24 hours later, I was just in Boxing Day. I was just completely shook cause I could see the, the, I could see the slope, right? It's slope on slope here, folks. It's, it's a derivative of the slope. Then February, I kind of really mechanic like made it very mechanical and that was the Z 80 cloning and [00:15:00] realizing what was going on.
[00:15:01] Geoffrey Huntley: That was like Sonnet 3.5. The way I described Sonnet 3.5 was it's. It's weird when you start really playing with the models, you start anthropomorphizing the models. And this is how I came up with the idea of quadrants of behavior, like high ethics, low ethics, an oracle, a deep thinker that doesn't take action.
[00:15:22] Geoffrey Huntley: And then highly, highly agentic. So I called, uh, Sonnet 3.5, a squirrel on cocaine with a chef's knife
[00:15:34] Andrew Zigler: you did. I remember that.
[00:15:35] Geoffrey Huntley: Yeah, it, it, it, I, I, I called it that because it, it could cook you a meal and it was so highly agentic, but it had a squirrel brain. If you took your eyes off it, it would murder your, it would murder your, your code base, and it would do it
[00:15:50] Andrew Zigler: to use every tool it could touch. It was tool hungry, tool happy. It was trigger, trigger friendly. It was a different kind of
[00:15:58] Geoffrey Huntley: It was a different type. [00:16:00] It really was this like hyperactive squirrel that would stab you if you took your eye off it. So you had different techniques to really control it, et cetera. And now like all, if you look at like cursed and all those things, and like why it looks like chaos. It's because of the damn squirrel.
[00:16:18] Geoffrey Huntley: It's 'cause of the damn squirrel. And now you don't need to do all that stuff because like Opus didn't suddenly get better over Christmas. Just people played.
[00:16:27] Andrew Zigler: Right.
[00:16:28] Geoffrey Huntley: and now you, it's, it's good. But the way I describe Opus is it's high, it's highly agentic. It produces great outcomes, but it's a little bit forgetful.
[00:16:39] Geoffrey Huntley: So you have to nudge it and remind it on, like, if you give it too many goals, it might forget a goal. So Ralph, and the idea of one loop, uh, just doing one thing means it's gonna be less forgetful. When GPT-5 first came out, it was in San Fran and we couldn't figure out what was going on because [00:17:00] we were, we were just in the office and we were like, we're tuning a harness.
[00:17:05] Geoffrey Huntley: And we're like, this thing feels like it's low testosterone. Like, 'cause there's this big webcast, et cetera, what's going on? And we're like, this thing feels low T and we couldn't figure out why. So went down across the road. Spoke with, uh, uh, essentially a competitor that's playing with this, and we're like, why is it like, why is it low testosterone?
[00:17:31] Geoffrey Huntley: And they go, oh, this is San Fran. You can't say low testosterone, Geoff. We prefer the words timid. And then, and then you start riffing. Like, what, what's going with the model? Like the, the, the, the, the harness community is really small. They might, might seem like enemies, but we're all trying to figure this out.
[00:17:48] Geoffrey Huntley: We anamorphize these behaviors and then we see what we can do to get rid. Once we've identified a commonality of behavior, what can we do to get rid of it? Now it turns out GPT-5 doesn't [00:18:00] like you using uppercase. If it considers it yelling at it and it gets timid. This is in the open AI model cards. Like if you yell at it. It, it, it beco it, it won't do its task. It actually does. So we stop yelling at it in our harness prompts and all our tooling. But this is where it gets really weird, because if you've got a code base with cursor rules and you've got a model, select a dropdown in and out, and you've done designed all your Cursor rules around Anthropic models which encourage yelling at the model that you want. They want you to hurl abuse at the model. But if you do it at to you, do the model selector, all of a sudden, is it really GPT-5 that's bad, or are you yelling at the model? Because all these Cursor rules now are yelling at the model.
[00:18:49] Geoffrey Huntley: Harness is now yelling at the model. And the official published guide on how to tune GPT-5 by OpenAI says, don't yell at the model. [00:19:00] Um,
[00:19:00] Andrew Zigler: Isn't it funny to think that they're not compatible on an API level, but then also on a behavioral level, the types of things you put in, the prompts on the, like the human created parts that go into it, or the text, BA based parts that go into it. Like it's even those they're, they're not compatible, they're not
[00:19:17] Geoffrey Huntley: I, I, my understanding is the MCP specification still doesn't have a user agent sniff topic, so it could switch out the, the prompts. Based on what the consumer is to take this type of nuance. And all the MCP servers you download and run, they're, they're the same story here. They've been built around anthropic, and then you use 'em in a different model and it's not so good.
[00:19:42] Geoffrey Huntley: It's not that model's no good, like it just hasn't been tuned.
[00:19:46] Andrew Zigler: That's right.
[00:19:48] Geoffrey Huntley: How do you know if you're tuning is because you start anamorphize the models and you start noticing these weird things and you come around other people who start thinking along these lines and you just try things. [00:20:00] It's a very different skill to building software, uh, with the tools, like tuning is a very different skill.
[00:20:06] Ben Lloyd Pearson: Yeah, there's definitely times that I've taken the same problem and taken it to a different model and gotten the result I wanted when I didn't get it from the, the first model. And, and I'm, and I'm kind of, I'm curious, you know, 'cause so, so Ralph is like, it, it's there, there's, there's two sorts of sides to this.
[00:20:21] Ben Lloyd Pearson: There's this like, brilliant simplicity that just takes a very simple concept and, and uses it to iteratively improve on, on an output. But it's also like. Incredibly inefficient, isn't it? Like is, is, is, is this how we're going to like tune these models or make them better? Or is it, are we making up for a failure of the capabilities that will get addressed at some point in the future?
[00:20:49] Geoffrey Huntley: The game is always changing. The game is always changing. Like the, the, the prompts I use for Curse and from February and all that stuff [00:21:00] for, to try to keep the squirrel on the, on the rails no longer apply. The game always changes. Like I still see people doing like GPT-3 Big. You are, you are, you are a QA developer, persona and all that stuff.
[00:21:15] Geoffrey Huntley: That stuff's dead. Folks go on. Gone. It doesn't matter. Like stop treating it like it's GPT-3, persona based is dead. Now to your question, doesn't mean you shouldn't adopt and start playing now just because the knowledge is disposable. Like you need to actually learn what it is and is not possible.
[00:21:36] Geoffrey Huntley: It's the capabilities and knowing on how you do loop backs and all this, and you learn all these tricks. The tricks, for me in the last year, I haven't lost. The approaches, how I get them have changed. The approaches, how I get them have changed. So, Ralph is incredibly inefficient because what you are doing is essentially mallocing an array, every [00:22:00] loop, and then telling you to only do one thing, and then you have to re-malloc this big thing at the top again, the specifications each time, each time.
[00:22:10] Geoffrey Huntley: So there is waste there. But if you think about that Ralph enables a path for founders to get software development at $10.42 an hour. Any types of talks of efficiencies is just gonna gonna drive that price down. So I mentioned the word array, and this is maybe some advice to, for developers in this space.
[00:22:33] Geoffrey Huntley: Context windows are tokens, are these weird MLE terms. I want people to think about this as an array and memory management. Mallocing an array. 'Cause that's all a context window is there's no memory on the server. API. There's none. What you are doing is you are, when you do a prompt, you are allocing or mallocing your prompt [00:23:00] to that array.
[00:23:01] Geoffrey Huntley: It literally gets sent off as a standard rest request. It gets churned on some GPUs and it might go, you advertise the capability that you could run a program, so that response comes back. Your harness says the LLM wants to run a program. Okay, I'm gonna run ls, and then it auto allocates the LS to the array.
[00:23:29] Geoffrey Huntley: Then it sends it back for more inferencing. You're just mallocing arrays here, folks. So, and that's one of the, been the biggest turn of inventions that, that caused from my o like oh god moment was I used to copy and paste code, uh, from GPT-3 into like an editor, press compile, then back forward copy and paste manually. All tool calling is the automation of that. Okay. [00:24:00] All Ralph is is the automation of tool calling for a system. In treating the entire thing as a system and taking systems based approach from first principles. So if you've got this array that you're mallocing and then there's some hints in the name of a context window, software engineers should know what a sliding window is. So you've got an array, you've got a sliding window. The window's only so big, the array is much bigger. So the less that you allocate, then the, then the more the window's gonna be able to see. So thus the, the thought that you should only pick one in the loop. And this is really wild because a lot of engineers right now are using, uh, multi-stage planning, and they, they, they're treating it like a child.
[00:24:44] Geoffrey Huntley: They, they, they, they're breaking it down to the steps they should do, and they're in this high control thing. Invert that thinking. LLMs are fantastic at prioritization. If you are building a brand new application up from [00:25:00] specs, it will do the logging module first. It will do the, the, the telemetry next. It will do repository patterns for databases.
[00:25:09] Geoffrey Huntley: It will, it will do the right things in the right order. So by just telling it to just pick the best. Like the most important item and only do one. You are, you're mallocing less into the context window because what happens with uh, a context window is you got an advertise number of 200k that's a marketing number folks, the same way you get a hard drive and it, it, it, like, it says it's uh, 20 terabytes, but it's only 18 terabytes usable.
[00:25:38] Geoffrey Huntley: Like the same thing here in thinking because there is a harness over overhead. Uh, there's a harness overhead, and then there's a model overhead. So a model overhead would be like the, the, the actual programming of, uh, Opus, for example. You can't get rid of that. So all of a sudden you can crudely say that you lose 16k [00:26:00] tokens to the model overhead, and then when you're using Cursor or any other Claude Code or what else have you, you also get a 16k overhead.
[00:26:10] Geoffrey Huntley: So the real number is like in the 176, but if you put in MCP servers in there, all of a sudden the real number gets down to like 120 and then you throw in some cursor rules in there. I've seen people operating with like, they've only, this is like a Commodore 64 worth of memory, and they, they've allocated over 80% of their memory before they even be able to get a loop going. And because this is a sliding window, right? It's so easy for it to get forgetful and all these other things. The more you allocate the, the more you, you're gonna get bad outcomes. So Ralph is a deliberate attempt to minimize allocation. So I never get a compaction event. Compaction is the equivalent of, imagine you got a, you got this beautiful context window, a beautiful array getting you really good outcomes, [00:27:00] and you, you, you, oh my God, there's been so you've seen them that like, you are like, I
[00:27:04] Andrew Zigler: No. And then it compacts like, Ugh, I'm so sad. Yeah. And then, and then it compacts and you're, you're,
[00:27:09] Ben Lloyd Pearson: or even if it doesn't compact it, it gets anxiety that it's going to have to compact it and then it starts making bad decisions
[00:27:15] Ben Lloyd Pearson: as
[00:27:15] Geoffrey Huntley: direct. So, uh, Sonnet 4 had the, was, uh, we just, we actually described it as when we're treating harnesses as it had context anxiety. Opus doesn't really have some of those attributes, but what would happen when it had context anxiety, it would go in a simple implementation. Uh, I'm running out of time, I'm just gonna like, comment out this thing and cheat.
[00:27:41] Geoffrey Huntley: Like, there's all these little meta things. So, so the idea of, I want people to think about compaction because it's a new term for folks, is really simple. Imagine that you've got this array and this array says this is your specifications, this is your goal. [00:28:00] Compaction is just literally imagine a Jenga tower, right?
[00:28:04] Geoffrey Huntley: And if you pull out the wrong bricks at the bottom, the, the tower falls over. Compaction is, is essentially this equivalence of it, like a Jenga tower and visualization of outcomes, but also if you download a video on YouTube and upload it a thousand times, download, upload, download, upload. It's actually a loss.
[00:28:21] Geoffrey Huntley: It's a lossy function folks. So what happens if the, the compaction removes the thing that was giving you this juicy golden window, which could be your specifications, and if the specifications get compacted out, next thing you know, the end result is it starts making stuff up. avoiding compaction and then using that array with a singular goal. To get stuff done. if we go away from Ralph really quickly, and let's just say someone as at Steve Yegge's, one or two. What I, a failure mode I see [00:29:00] is software developers. They click new chat, they get the, they, they like make my website pink, and then they reuse the array and they set another goal that's completely nothing to do with the website being pink.
[00:29:12] Geoffrey Huntley: And then they say, make me an API. All of a sudden they've got this backend API control with pink rest endpoints like you, you need to be thinking about a new chat is a new array. Each array needs to be chat hygiene. And before we go past that, I'll just rewind a bit. There is proven scientific research of the notion of context rot and the existence of a dumb zone.
[00:29:41] Geoffrey Huntley: So if you think that you've got like a 200k and like once, once you go like 60, 70% over that, like the LLM starts getting dumber, actually dumber. So you want to like minimize the amount of time in your dumb zone is what, uh, Dex Haughty would say. And it, it is [00:30:00] true. And so Ralph deliberately is. You don't want to get in the dumb zone and you don't want a compaction, but you want to deterministically allocate.
[00:30:11] Geoffrey Huntley: It's an orchestrator pattern for, for a Gas Town or for what is to come next or someone can use to build their own gas town.
[00:30:21] Andrew Zigler: That's right. You know. Let, let's, let's talk for a minute about, about Gas Town. You bring up Steve Yegge and, and, and maybe how Gas Town is, uh, something that's building on the idea of Ralph maybe is something totally different from Ralph. I'm curious to know from your perspective, like what you're teaching all of us to do right now and embracing.
[00:30:38] Andrew Zigler: The slope on the slope. There's an exponential curve of learning and opportunity here for folks. And I think the real takeaway from this is that the simplicity of this loop, the simplicity of all of these parts that you have to bring in to keep context small, to be light weight, and to be interchangeable, I think all of that plays into people building their own types of Ralphs or own loops.
[00:30:58] Andrew Zigler: And maybe [00:31:00] Gas Town is an example of one of those coming into fruition. But you know, I. I personally as someone, uh, who's building and using these tools too, I see it as an inspiration to go and maybe craft my own gas town. And I've definitely borrowed a lot of concepts from it. So I'm curious like how you think about the evolutions of your concept like Ralph and how you see them out in the world.
[00:31:20] Geoffrey Huntley: Yeah. I used to work with Yegge over at Sourcegraph. I spent a bit of time traveling around APAC telling people. Hey, please pay attention. Pay attention. Pay attention. 'cause I was sitting on Ralph and Yegge was out doing much more worldwide big tours 'cause he is Yegge and he's just saying, folks pay attention.
[00:31:42] Geoffrey Huntley: The orchestrator is coming. And he'd already done like his rant saying this is gonna happen. And I don't think people took him seriously. So on New Year's Eve, he dropped the blog post, said, welcome to Gas Town. Do not use Gas Town. Gas Town is not for you. But by the way, [00:32:00] here's this beautiful chart that allows you to measure where you are on skill level and in the figure five is the notion of where you start deterministically allocating a agent and Ralphing and
[00:32:18] Geoffrey Huntley: when you get to like six and seven, six and seven, six is when you are stating to like learn to run model these at the same time. And then you experience failure domains when you're running two at the same time and then you realize it's, you've gotta do some engineering. So like you, the two spinning plates you've got going don't fail in the same way. By the time you get to like figure seven, you've got like 10 plates spinning at the same time. It's just different windows on a wide screen monitor. And it starts feeling really chaotic. It feels like a spaghetti base in factorial. It feels like a spaghetti base in factorial. Where did I put my iron?
[00:32:57] Geoffrey Huntley: Whatever I my oil. And
[00:32:58] Andrew Zigler: Where is it? Where is it? [00:33:00] It's all in arm's reach, but where the hell is it?
[00:33:02] Geoffrey Huntley: Where the hell is it? Right? And the reason you shouldn't go to Gas Town is it's, it's Like use Gas Town is you need to go through the motions of a, per yourself development as a, as a developer to learn why Gas Town needs to exist and why it's coming.
[00:33:19] Geoffrey Huntley: And that's from like learning to allocate the array in figure five. And then also changing your software engineering practices. 'cause by the time you're doing figure five, which is Ralph, like a singular. What happens is you are, you start questioning things like code review. Why do we do code review? I think that which you get that you start thinking like, why do we have agile?
[00:33:43] Geoffrey Huntley: Why do we do daily standups? It's our job as engineers to potentially falsify the last 40 years of software engineering practices and engineering magnet management practices for our new power tools that now exist. [00:34:00] Everything is now different. But we're still an engineer. We don't want failures. We don't want people like things getting hacked and all the rest.
[00:34:08] Geoffrey Huntley: So like at figure five, like, you, you're generating a 40,000 line pull request in a couple hours and your coworker's doing the same, and like you, it starts to be transparent in your face. Like I think the industry's gonna spend the next year going, how do we fix the code review dilemma? The answer is you don't for classes, depending on the domain, especially in product, you just don't do the code review.
[00:34:35] Geoffrey Huntley: Instead, you, uh, you, you work on safe release practices. You engineer, if you see value domains such, you put 'em as pre-commit hooks. Maybe you run other Ralph loops. After the Ralph Loop is done. That's when you start thinking into, uh, when you start getting the megabytes in factorial. And that's Gas Town. And that's for [00:35:00] the lessons learned at 5, 6, 7, 8 is the full application of all that learning that came before. You can't just skip the learning and just go straight to Gas Town. Like maybe you can, but you're not gonna have any edge. Like, uh, we're meant to be curious as engineers go for and earn your stripes.
[00:35:22] Geoffrey Huntley: It's not that, not that complicated. Now at seven, you seven, you, you create eight. Because it's just too chaotic. It's so draining. Like trying to find resources in your, in your base be anytown is just so damn ridiculous. You like stuff it, I'm gonna rebuild my base from all the knowledge and then you, whatever your domain is.
[00:35:44] Geoffrey Huntley: If you're a software engineer, you're an accountant, you are a, a, like legal, whatever, you will build your own gas account. I've, I've got my own gas town. Uh, it's called Loom. cloned GitHub, uh, [00:36:00] amp code, uh, Daytona. I have my own source code called jj. I have the ability of a source code host, like I, I'm basically on the idea if I want to explore the, the space of invalidation. I essentially need to be able to control everything from source to be able to change what it is, and now it's so simple to clone product features and companies. The next thing I'll do is like Launch Darkly. I'll have feature flags in there so I can play around with safe engineer. You can literally clone platforms now folks, the models are getting so good.
[00:36:38] Geoffrey Huntley: It's getting so simple. So I have my own gas town, but my gas town is like a hard fork. From, uh, everything that's been done today, I'm, I want to build an autonomous evolutionary software loom. Yegge is within the realms of what exists today. [00:37:00] That's a differentiator. Like I, I, I'm prepared to like fork like source control.
[00:37:07] Geoffrey Huntley: I've got my own file system that I've built for the agents. Like if you think it's unhinged, I'm doing it. Trying to find the answer of what the right thing to build and not build is, and what do agents need? Like,
[00:37:22] Andrew Zigler: I don't think it's unhinged. I'm inspired. It's.
[00:37:25] Geoffrey Huntley: I'll, I'll give you something to think about. Uh, Unix is essentially 40 years of design for humans.
[00:37:34] Geoffrey Huntley: Think about it.
[00:37:35] Andrew Zigler: Mm,
[00:37:36] Geoffrey Huntley: we, we have the TTY, which was on the idea of the humans interacting it and we got bash 'cause it was for the humans. We got a shell for humans
[00:37:45] Andrew Zigler: all programming languages, just evolutions of a base language for the computer. But
[00:37:51] Geoffrey Huntley: environment
[00:37:52] Andrew Zigler: All for humans. Environmental variables for humans.
[00:37:54] Geoffrey Huntley: Why do we, if humans, for humans, for humans, so humans, and then you [00:38:00] start thinking. Wow. What happens if I start cutting down the stack?
[00:38:04] Geoffrey Huntley: Like I'm, I'm, I'm familiar with like Unikernel domain, knowledge of research, et cetera, but I'm like, what if user space didn't exist anymore? That's where I'm going to explore, like, what is agent space? What if we took the last 40 years of Unix design and we made figured out what agent space, what if in.
[00:38:27] Geoffrey Huntley: What if it's just SIS calls to a kernel. And that's all an agent needs? Like we need to rethink about what engineering is. We've got this magic like. We've got this magic space dust machine that can do crazy things. And I guess when people first do their Ralph loops, they're gonna feel like, I did you feel like you're robbing your future, but not in the sense that you, you, you, you're putting yourself out of a job.
[00:38:59] Geoffrey Huntley: I've done all [00:39:00] the things I wanted to do in retirement. It's, it's, idea to execution. Like I wanna build a file system. Oh yeah. I built one last year. I wanna able to build a program. Oh yeah. I built one last year. Like you start doing all, it's, it's like, fuck. And then you realize there's this mass like dopamine that you can just do anything now and you start just doing it.
[00:39:19] Geoffrey Huntley: And we're gonna see this mass explosion of creativity of people just doing it. It's gonna be like geo cities, it's gonna be really pretty, it's gonna be scary for founders. Go speak with founders how they feel about this because you know, it is pretty unhinged that one person in three days just live streaming.
[00:39:38] Geoffrey Huntley: Uh, essentially. I don't know, three, four companies of a thousand employees each over six years worth of work that they would've put in in those three companies and got the core feature set in three days. This is what we're looking at here, folks. But this is when I said the Rift is getting massive. And if people [00:40:00] were still sitting there with cursor going, why no?
[00:40:02] Geoffrey Huntley: Why no. Know that there are people out there doing laps. They're, they're doing laps on laps right now.
[00:40:11] Ben Lloyd Pearson: And something that Andrew and I have actually talked about, uh, just on our own over the last year is how, like it does feel like it's never been easier to get ahead of everything. As quickly as possible, like you can just go, you, you, you learn a new technology in this space, and then suddenly you're operating a hundred times faster than anyone else around you.
[00:40:31] Ben Lloyd Pearson: But it's also never been easier for everyone else than to catch up and figure it out and get right back to where you are. You know? I mean, Ralph has been a thing for, for, you know. Uh,
[00:40:40] Geoffrey Huntley: Six months.
[00:40:41] Ben Lloyd Pearson: six months or
[00:40:42] Geoffrey Huntley: Yeah. Seven.
[00:40:43] Ben Lloyd Pearson: then suddenly everyone wakes up and is like, oh my gosh, I should also be doing this thing.
[00:40:48] Ben Lloyd Pearson: And now suddenly everyone like, sort of is, and I, and I, and I hate this phrase 'cause it gets way overplayed, but like a lot of people are saying like, AI won't replace you. Someone using AI will replace you. And [00:41:00] I'm, I'm almost kind of getting the sense out of this that like, you know, gas town isn't going to replace engineers, but engineers who are using Gas town or who, who are building their own gas town, those are the ones that are going to like replace the, the other engineers, and so in that, and when you frame it that way, it's like learning how to build a gas town and how to do this Ralph technique.
[00:41:21] Ben Lloyd Pearson: This is actually just the new domain of, of fundamental skills that. Engineers need to have. And it's, you know, just like learning a programming language, it's something that you, you have to, you have to build that skill and refine it.
[00:41:35] Geoffrey Huntley: Uh, um, this is the hard I, this is the hardest send in agreement. We understand. We've got a birth of a new discipline of compsci here. Like it, it's been invented every day, every every day.
[00:41:50] Geoffrey Huntley: If you're not hanging around people who don't see this, then you need to plug yourself into people like, uh, there's some good communities on Bluesky, but most of the AI [00:42:00] activity for devs is happening on X.
[00:42:02] Geoffrey Huntley: Like, go plug yourself in and be really curious. So one of the first fundamentals every developer should learn is how to build your first coding agent. 'cause a coding agent is your tool. We spend hours and hours and days configuring Neo VM and all our dot files, et cetera. If it's just 300 lines of code, why would you not want to have your own coding harness?
[00:42:30] Geoffrey Huntley: Then you can exercise your taste, your discipline, all that high control you used to do on like on code. You can now do it on your printing press of code. So go build your agent because like Cursor, Windsurf. All these like harness companies are building coding agents. There's nothing different between them.
[00:42:52] Geoffrey Huntley: There's pretty much nothing different between them. It's 300 lines of code allocating array with some tools like read [00:43:00] file read file, edit file, bash tool, and a search tool. A search tool is just executing the bash tool and executing the command rip grab, that's all they're doing, and they're trying to differentiate themselves as being like. Like the Gucci or the Louis Vuitton and we were different, or, or we, we, we, we, we, we make
[00:43:21] Andrew Zigler: I'm a cursor user. I'm a windsurf user. I like camps. Putting everyone into camps. But you don't have to be in a camp.
[00:43:28] Geoffrey Huntley: you don't have to be in a camp. Like honestly, I trust a, a car mechanic that's rebuild a few carburettors a whole lot more than someone just orders the new part and slaps it on. I
[00:43:41] Andrew Zigler: right.
[00:43:42] Geoffrey Huntley: Software, software engineers that are curious, that are being mechanics, they've go rebuilt the carburetor. They understand. Exactly. 'cause once you understand what curse is doing, you go, you understand the prestige, you realize it's not magic, it's not [00:44:00] scary. It's like I'm scared of a wild true loop that with rest API.
[00:44:04] Geoffrey Huntley: Yeah, you are. you realize just how, how dumb that is.
[00:44:09] Andrew Zigler: Right.
[00:44:10] Geoffrey Huntley: and then you go build your own, and then you, you start thinking, well, okay, I'm gonna automate workflows in my life. I'm going to automate like a producing of software, and then you eventually end up at Gas Town. So I spent. The last year really encouraging people to go build your own agent.
[00:44:27] Geoffrey Huntley: I've got a workshop. It's on GitHub. It's free, it take you 30 minutes and it, it, it, it's unlocking people's brains. Now where that gets interesting is because once you realize the magic trick of AI and why every company is doing AI is because it's just those 300 lines of code with like a prompt on top GPT wrapper stuff. You realize there's no moats. So
[00:44:57] Andrew Zigler: Mm.
[00:44:57] Geoffrey Huntley: this is something that Dev should have a [00:45:00] conversation with founders and business leaders. What is a moat now? If some person in Sydney is able to basically rip a far into a voice prompt and like clone huge product features in companies that would've taken thousand, 2000 over 10 years. Just not do it at one and do it at many. And if I come into the market type topics or someone comes into the market, if there's like 10,000, uh, employees or heads to be fed like, like outgoing expenses and at a singular company, and then it's just three chill dudes in Bali, like living like kings, and they coming in at a different unit
[00:45:47] Geoffrey Huntley: economics, guess what's gonna happen? So, uh, we started on this story arc of like, Like AI is gonna take your job, et cetera. No, this was, I was, this is one [00:46:00] of my things, like some software developers not gonna make it, and I framed it really around the idea of employee performance bell curves. Some companies do performance reviews at six months, some do it 12 months high performance, six months, it's very aggressive, FANG, et cetera.
[00:46:17] Geoffrey Huntley: And what's happening is you're gonna, you come across FANG et cetera. Let's say that you're a high performer at FANG like last year. This year, when you do performance reviews, your, the high performance is now low performance because the high performers are Gas Towning or they built their own gas town and that's why you won't make it. It's not ai. You just, you weren't curious, you didn't invest in yourself. You were asleep at the wheel and the inevitable happened. And business leaders need to actually make adjustments because there's a couple people having, a couple people having the time of their lives coming into their market, a cheaper price point, and they're never gonna raise money.
[00:46:57] Geoffrey Huntley: They've had to raise money, they've got valuations, got [00:47:00] stakeholders and investors. This is Clayton Christianson disruptive innovation. Textbook. So if you want to be a developer, don't be, don't be, don't be a waiter, don't be a Jira ticket monkey. Be a doctor that diagnoses the problems. So like, you want to get yourself as close as possible to a PM or in front of customers, and you want to like. You don't want to be downstream because when you're at figure five Agile fractures at figure eight, like you're like, yeah, I'm just gonna be on Bali and I've got my Loom that's automating software development and I can do anything I like. I just gotta like throw some tokens on the campfire. And this is the evolution of where software is.
[00:47:52] Geoffrey Huntley: We're still engineering, we're still software engineers. It is critically important to be around people who get it. [00:48:00] And I would like to add the following note. Just because AI is not working in a company does not mean AI does not work. It means the company has a problem with AI because they've got a whole bunch of corporate dogma.
[00:48:16] Geoffrey Huntley: Maybe they're doing have to do a lot of context engineering to try and get it to the right coding standards and patterns. But the reason we had coding standards and patterns was for legibility for humans, getting back to the full humans thing. So I've seen people trying to, like they, in woodworking, there's a notion of working with the grain or against the grain.
[00:48:39] Geoffrey Huntley: I see a lot of corporates working against the grain and then the employees using AI there, they're not using it at home and experiencing working with the grain. And uh, my biggest concern is those people. Don't. When someone says, hi, I am a software developer. If they say, AI doesn't work for me, [00:49:00] I go, sweet.
[00:49:01] Geoffrey Huntley: What identity are you coming from? Are you coming from as a software developer where your identity is that you, you're a .NET developer? Well, 'cause that doesn't matter anymore. Are you coming from that? You're, you're coming from like Deutsche Bank? Well, that, that really doesn't matter anymore. Like if AI is not working at Deutsche Bank, that's a Deutsche Bank problem.
[00:49:19] Geoffrey Huntley: I'm speaking to the developers like, are you playing with it at home? Are you engineering and learning the way? Because we've got a year now. We've got model first companies like working with the wood, where they've captured the back pressure. They've, they've engineered ways so Ralph can keep on working.
[00:49:40] Geoffrey Huntley: Understand there's a completely different mental model on how software should be built now, different unit economics, and if your company doesn't get it. You have a choice. You can either go to a company that gets it there, there are plenty out there now. You can go capture the opportunity and live like a king in [00:50:00] Bali. Uh, and build your own gas town and just like smash your previous employer. Or you can suffer through a probably three or four year AI transformation program if any developer remembers when all the Essentia consultants will run through. And we got Agile coaches. That's coming, baby,
[00:50:19] Ben Lloyd Pearson: I think it's already here for a lot of
[00:50:21] Andrew Zigler: Yeah, I'm pretty sure that's already like more or less happening a lot of places.
[00:50:25] Andrew Zigler: Like I'm just, I have so much I wanna follow up on, but at the same time, we've been chatting now for a bit, so I wanna be mindful of time Geoffrey and this conversation, we've taken a wild tour through loops and into towns way out in the distance, and you've really kind of painted a picture about where this is all going and, and that's why we were really excited to tune in today because, you know, on Dev Interrupted, we've constantly, you know, been following the stuff that you've been talking about, how the moats are gone, how the walls are gone, how people need to be running hill climbing and, and building their own Ralphs.
[00:50:57] Andrew Zigler: And it's just been an amazing opportunity to kind [00:51:00] of dig deeper into where you're thinking about all this, especially capitalizing on this amazing zeitgeist of activity happening right now post-holiday break, where everyone's emerging, freshly experimented with these tools. Maybe they had put it down years ago and then pick it back up, but now they did.
[00:51:16] Andrew Zigler: Like there's an elucidation in the air and you can smell it when you talk to people about how they're talking about how they're working in engineering right now. And so we're gonna keep having these conversations. You know, you're a friend of the show. Uh, we want to continue understanding where this is all going and I think you're the perfect person to guide us in our audience.
[00:51:36] Andrew Zigler: And so just, and you know, keep that in mind that we're gonna be having you back here in the hot seat very soon. But before, uh, we do start to close out here, I just wanna turn it back to you. 'Cause everything. All, all good things are a loop, Geoff, so I'm gonna turn it back to you one, one last time and say, where do you wanna leave this on?
[00:51:54] Andrew Zigler: Where should folks go, uh, to go learn more about you and what you're building and where this is, uh, how [00:52:00] they can start, get, getting advantage of this themselves.
[00:52:03] Geoffrey Huntley: Yeah, sure. Um, my website, uh, ghuntley.com. You'll link it in the show notes. Uh, it's all there. It's all free. Just an email address. That's my ask. If you want to pay, you'll get access to. Some of my ideas, maybe 48 hours early. San Fran crowd likes the notion of alpha. They like to front run and like have the ideas first.
[00:52:26] Geoffrey Huntley: So if they wanna pay for that, you can pay for it. But you don't have to. You just like put your email and you can see it all. Been publishing for the last year. Uh, it's all about the loops. Now, when I say engineering, you should be engineering feedback loops. So that's all Ralph is, is like forcing the feedback loop back on itself to actually get something to done.
[00:52:46] Geoffrey Huntley: And that's all that tool calling was. And it's your job now to actually start looking at data sources that you can engineer to feed back into the loop, so to, to get the outcomes. Now, if this is all [00:53:00] new to you and it's really startling like that, there's people a year ahead and they're doing absurd things, you're like, I can never catch up.
[00:53:07] Geoffrey Huntley: That's bullshit. Would you believe if you're watching this, you're early. Only 10,000 people have installed the cord code plugin. You're early. There's still time. There's still time. Strap yourself in. Just know that the official Clade Code plugin, uh, does, it takes you on a path. Where you, uh, where you get compaction, et cetera.
[00:53:33] Geoffrey Huntley: They did a really good job of accessibility, but you need to approach this like an engineer. Just go, go do it. Like build it up step by step. Like Kubernetes is the hard way, but like context engineering. I have another video where I with Dex Horthy on my YouTube that says, why the philanthropic plugin, isn't it?
[00:53:54] Geoffrey Huntley: And maybe you should go look at that as well. It gets in into really a detail of context engineering and [00:54:00] some other tips. Go check it out, and I can't wait to come back.
[00:54:03] Andrew Zigler: Amazing. Well, thank you again, Geoff. So happy we made it to the end of the call before a compaction event occurred. Even though I think that between our conversation, we definitely filled the whole window with a lot of ideas for people to be digesting. So folks, you know, definitely be processing that we're all gonna be coming back in this hot seat and talking about this more soon.
[00:54:22] Andrew Zigler: Uh, 2026 is gonna be a fascinating year. We're orchestrators emerge and, and, and you and your team, they're already full of them, so let's take advantage of that opportunity. Geoff, thank you again for coming. Uh, and, and, uh, we'll see y'all at the next one.
[00:54:35] Geoffrey Huntley: To the next one. See you soon.
