Inventing the Ralph Wiggum Loop | Creator Geoffrey Huntley

0:06Hey there, and welcome back to Devinterrupted. I'm your host, Andrew Ziegler. And I'm your host, Ben Lloyd Pearson. And I'm excited to announce that we're doing something a little different this week and starting this week, actually, and moving forward. First of all, all these long-form interviews that we publish every Tuesday, they're not going anywhere. We have a really wonderful guest, multiple guests, lined up for upcoming episodes. And we've got another really fantastic one today that we'll get into in just a moment. However, we are going to split off our normal news segment into its own show every week.

0:38that'll come out on Friday. So this Friday will be the very first iteration of this new show. So tune in this Friday to get news updates from us. We'll even try to bring in some experts from the field to share really cool things about what's happening in the world of AI and engineering productivity. Those are the changes. Andrew, what do we have coming up today? If you're an engineer and you've been online at some point in the last week, there's a good chance you've seen Ralph Wiggum take the industry by storm. And what is Ralph? In its creator's own words, Ralph is a bash loop. And that creator is Jeffrey Huntley.

1:11Jeffrey has been pushing one of the simplest, most provocative ideas in autonomous AI-assisted coding. And Ralph, it turns any iterative work into something that keeps fixing, refining, and reevaluating until tests pass or succeed. And really the brute force effort of this can't be understated. The amount of work that Huntley can do with this type of loop is truly profound. And this work, it sits at the intersection of practical workflows, philosophical insights, a little bit of contrarianism, and a very bright glimpse about how the future is going to be built with simple loops, clear goals, and letting feedback drive the convergence towards done.

1:52And we've been covering Jeffrey's work here on Dev Interrupted for the past year. Today, he's joining us to dive into Ralph's origin story and what that means for you now that the cat is out of the bag. Yeah, and we discussed quite a few deep topics. We'll have links to all this stuff in the show notes because there's a lot of reading that I think goes along with this episode. And one of the concepts we'll talk about is this new article from Steve Yege called Gastown, which just describes a very complex system of effectively applying the Ralph concept to much larger challenges. So I think this is one of the most interesting discussions we've ever had at Dev Interrupted.

2:32And there's a lot of things for our audience to dig into on this one. And of course, all of this content will be our newsletter and Substack and LinkedIn. So be sure to find it, join the conversation, as well as your fellow developers as you figure out what all of this means for us. Now let's dive into the interview. This is so exciting. I'm actually kind of excited that we also have those lost tapes because that was some really good shit that you were putting out there a second ago. Those are our secrets now. Yeah, and now those are like, wow, those are now only I heard that. It was a closed room.

3:05That was great. But so this is just like overwhelmingly cool, Jeffrey, like to have you here. 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 Curser to you. All of what I've learned in the last year started with reading your articles as like the nucleus of that thought. 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. And because I watched you bring the now very well-known Ralph into the world.

3:43And that's what we want to talk about today. and what it means for developers that are maybe using their own Ralphs. Because I think there's 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 out in the field that are experimenting. So we're really stoked to have you here. But I want to just also frame the Bash loop itself and kind of take a step back because I want to start there and just ask you, when you brought Ralph into the world and you named it Ralph, why give it that name?

4:22Yeah, so it's February last year, almost a year ago. And I was doing spec-based development. I made a mine at GitHub Next, published actually some research there about two years prior to that. And I remembered that on like spec based workflows or LLMs generating apps. This is like the GPT-3 days. So I started applying it and then I started realizing that like I could probably, this could be quite mechanical. Like if I take specs and how I'm doing things. And then I got to the point where I was really like, wow, there's some pure first principles thinking here of like, I could deterministically allocate the array.

5:12And the first thing I was like, okay, what can I do with this? And this was the Z80 blog post. And I started the path of essentially using LLMs to essentially clone product features and of existing companies. Yes. And like taking source code that was like BSL licensed source code and like clean rooming it into specs, very similar to how the Intel and AMD CPU came into. And I re-implemented like HashiCorp Nomad. I cloned the company Tailscale. I did all these random things. I remember Rovo. We read Rovo. Rovo. You did several of them. It was really interesting to watch how you use the LLMs to do this binary analysis, and then you would pull out this information and rebuild what they had put in it.

6:13And it really kind of took away... It drained any moat that you could find. 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. Pretty much. So for developers, you go speak with other founders and how they're feeling about this. Because it's pretty gloom out there for founders. Like we can focus on the developers and developers side of things. But go speak with your founders. Now, to answer your question, when I discovered this in February, it literally made me want to Ralph.

6:53Ralph is a term for vomiting. And it wasn't named. It wasn't named. Like it wasn't a technique was not named, but it made me want to vomit because I could actually see where we're going. I could see where we're going. Like I was building software in my sleep and then I started publishing a lot more. 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 I really focused on some of you not going to make it. You need to be curious.

7:35You 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. that if you drove it in the right way, you allocated the context window deterministically in a loop and it was all so simple that I could see where it could go. 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, they just don't know that they don't have a job yet and I'm like, wake up, wake up, please wake up, please be curious.

8:15like play with the llms 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 ran into dex haughty 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 meet up in San Fran two hours late, presented last, and we just kept going all night. And it was like 15 people in there. We spent an hour afterwards going, oh my God, they could see it. They could see it. And because it's San Fran, I decided, hey, look, the word's going to get out about this thing because I just essentially showed the technique for the first time.

9:10and 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 want to Ralph, so I called it Ralph, as in Ralph Wiggum. And thus, it forever is now stuck in meme folklore. 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 is dead, but software engineering is very much alive, okay? If your value to a business is that you type things on the keyboard, would you believe Sonnet 4.5 on a loop with a bash loop, Ralph?

10:11It costs$10.42 US an hour. Now, what's the minimum wage laws in your state? I'm pretty sure that's why the discussion in Sanfram 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 the unit economics changed. It changed six months ago. There's been a batch of Y Combinator companies, Ralphing, you folks were early, you're reading, you're growing. There now exists people who are a year in, or six months or a year in have been applying it. And the rift on this is becoming huge. huge huge so so unfathomably huge it is unbelievable the difference someone who's been curious for 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 engineering however is alive software engineering is alive if you if you're 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 an autonomous worker to do your job now software engineering is a different discipline this is the the like actually engineering if the idea of autonomous loop within your code base makes you want to ralph listen to that feeling right 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 cents to the ton after the shipping container the ralph loop is that same type of equivalence of the change in the unit economics.

12:23It's not going back. So if you're the software dev or the sailor, like carrying cargo on and off the ship manually, please understand the container, the box just got invented. Now, engineering is different to software development and engineering is now the game. Because let's say running an autonomous loop, If you're like, oh, what happens when it drops the database? Sweet. You're an engineer, right? What are you going to do to stop it from dropping the database? Don't provision right secrets. Maybe you introduce some tests and enable change data capture. And then you expect the audit log. Like, engineer your way out of the failure scenarios.

13:10And the more you do that, the more you capture the back pressure, which allows more autonomy for the new unit economics. That's the game now. I'm like pre-commit hooks, property-based tests, snapshot tests. And eventually you'll come to where I am, where I don't even use CI. And my agents run full with sudo on a machine, bare metal box running Nix OS. and I don't review the code, the agents autonomously push to master, no branches, and the deployment happens in under 30 seconds. And then if something goes wrong, I've got feedback loops, which feed back into the active session and then it just self-repairs.

14:07So we're starting to get into it. When you start to get into some what I'm researching, I call this Loom and Steve Yeager calls this Gastown. We're exploring what it means to be self-evolutionary software. 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 connect us to some other data source and then they make the cutover and they roll it all out again. So this is engineering now, folks. It's all an engineering game now. You know, you're talking early on about the importance of that curiosity, you know, and I agree with you.

14:52And 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 know where to look. And I feel like a lot of people, their perception of this technology is a little bit tarnished by what may have been early experiences with it. Like, you know, we've all been using these tools since the days of GPT 3.5 or whatever it was. And I know that the things that it failed on back then were obvious often. Often I felt like it really did not seem like a very effective tool. But, you know, here we are. We've had multiple generations of models come out.

15:31The models have gotten substantially better in the last year in particular, I think. And I feel like there's – 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 of this all looks like. Yeah. Yeah, I've got some thoughts here. I'm actually kind of curious if you felt something similar, even with these autonomous loops that you're building, that you're ralping, are you seeing the changes to the models even resonate into improvements for that method of doing things as well? 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 like 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 wasn't like regurgitating stack overflow that was my mindset and my frame I'll just regurgitate stack overflow I tried this in G53 it's not anything and I came back because I ran it in a loop like by hand in cursor I wasn't quite yet at the very like novel idea of just running in a bash loop.

16:52It was just very mechanical. Like I was just doing the motions of specs and everything else.

17:00And that just shook me. And then when I got to February, this was at GPT-3. So that shook me. My previous frame of reference was like GPT-3-ish. So that was like, it took a call with a founder to say, Jeff, you need to get on the tools. like you need to start writing. Do you see what I see? I'm going to hang up now and you go play and you call me back. 24 hours later, I was just, in Boxing Day, I was just completely shook because I could see the slope, right? It's slope on slope here, folks. It's a derivative of the slope. And then February, I kind of really mechanical, made it very mechanical and that was the Z80 cloning and realizing what was going on.

17:50That was like Sonnets 3.5. The way I described Sonnets 3.5 was, it's weird. When you start really playing with the models, you start anamorphizing 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, and then highly, highly agentic. so I called Sonnet 3.5 a squirrel on cocaine with a chef's knife you did, I remember that yeah, I called it that because 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 code base and it would do it all the time it could touch, it was too hungry tool happy it was trigger trigger friendly it was a different kind of model it was a different type 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 etc 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 it's because the dance grill and now you don't need to do all that stuff because like opus didn't suddenly get better over christmas just people played um 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 so you have to nudge it and remind it on like if you give it too many goals it might forget the goal So Ralph and the idea of one loop, just doing one thing, means it's going to be less forgetful.

19:41When GPT-5 first came out, I was in San Fran, and we couldn't figure out what was going on because we were just in the office and we were tuning a harness and we're like, this thing feels like it's low testosterone. Because there's this big webcast, etc. What's going on? And we're like, this thing feels low T. and we couldn't figure out why. So went down, crossed the road, spoke with essentially a competitor that's playing with this, and we're like, why is it low testosterone? And they go, oh, this is San Fran. You can't say low testosterone, Jeff. We prefer that word's timid. And then you start riffing.

20:30What's going on with the model? Like the harness community is really small. They might seem like enemies, but we're all trying to figure this out. And 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 like you using uppercase. It considers it yelling at it and it gets timid. This is in the OpenAI model cards. like if you yell at 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 selector drop down in and out and you've done all your cursor rules around anthropic models which encourage yelling at the model that they want you to hurl abuse at the model.

21:26But if 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. The 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. 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'd put in the prompts, like the human-created parts that go into it or the text-based parts that go into it, even those, they're not compatible. They're not compatible.

22:07My understanding is the MCP specification still doesn't have a user-agent sniff topic. So it could switch out 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 the same story here. They've been built around anthropic and then you use them in a different model and it's not so good. It's not that model's no good. It just hasn't been tuned. That's right. And how do you know if you're tuning? It's because you start anamorphizing the models and you start noticing these weird things and you go around other people who start thinking along these lines and you just try things.

22:48It's a very different skill to building. software with the tools. Like tuning is a very different skill. 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 first model. And I'm kind of, I'm curious, you know, because so Ralph is like, there's two sorts of sides to this. There's this like brilliant simplicity that just takes a very simple concept and uses it to iteratively improve on an output. it but it's also like incredibly inefficient isn't it like it 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 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 for to try to keep the squirrel on the on rails no longer apply the game always changes like i still see people doing like gpt3 big you are you are you are a qa developer persona and all that stuff that stuff's dead folks go on go on it doesn't matter like stop treating it like it's gipty 3 persona based is dead now to your question that 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 it's the capabilities and knowing on how you do loopbacks and all this and you learn all these tricks the tricks for me in the last year i haven't lost but the approach is how i get them have changed the approach is how i get them have changed so So Ralph is incredibly inefficient because what you're doing is essentially mallocing an array, every loop, and then telling you to only do one thing.

24:52And then you have to re-malloc this big thing at the top again, the specifications each time, each time. So there is waste there. But if you think about that, RALF enables a path for founders to get software development at$10.42 an hour, any types of talks of efficiencies is just going to drive that price down. So I mentioned the word array, and this is maybe some advice for developers in this space. Context windows are tokens of these weird MLE terms. I want people to think about this as an array. and memory management, mallocing an array. Because that's all a context window is. There's no memory on the server API.

25:41There's none. What you're doing is when you do a prompt, you're mallocing your prompt to that array. It literally gets sent off as a standard REST request and 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. Johanna says, the LLM wants to run a program. Okay, I'm going to run LS. And then it auto-allocates the LS to the array, and then it sends it back for more inferencing. You're just mallocing arrays here, folks. So, and that's one of the biggest turn of inventions that I caused from my oh god moment was I used to copy and paste code from GPT-3 into like an editor, press compile, then back, forward, copy and paste manually.

26:43All tool calling is the automation of that. Okay. All Ralph is the automation of tool calling for a system and treating the entire thing as a system and taking a 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 is only so big, the array is much bigger. so the less that you allocate then the more the window is going to be able to see so thus 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 multi-stage planning and they're treating it like a child 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're building a brand new application up from specs, it will do the logging module first, it will do the telemetry next, it will do repository patterns for databases, it will do the right things in the right order.

28:01So by just telling it to just pick the best, like the most important item and only do one, you're malicying less into the context window. Because what happens with a context window is you've got an advertised number of 200K. That's a marketing number, folks. The same way you get a hard drive and it says it's 20 terabytes, but it's only 18 terabytes usable. The same thing here in thinking because there's a harness overhead. There's a harness overhead and then there's a model overhead. So a model overhead would be like the actual programming of Opus, for example. You can't get rid of that. So all of a sudden you can crudely say that you lose 16K tokens to the model overhead.

28:51And then when you're using cursor or any other code code or what else have you, you also get a 16K overhead. 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 more you're going to 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 ray getting you really good outcomes and you oh my god there's been so you've seen them that like i know and then it compacts like oh i'm so sad yeah and then and then it compacts and you're 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 as a correct so that is sonnet four 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 i'm running out of time i'm just going to comment out this thing and cheat.

30:31There's all these little meta things. 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. Compaction is just literally, imagine a Jenga tower. And if you pull out the wrong bricks at the bottom, the tower falls over. Compaction 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.

31:10It's a lossy function, folks. Right. So what happens if 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 so avoiding compaction and then using that array with a singular goal to get stuff done now if we go away from raf really quickly and let's just say someone as that steve yeggies one or two what a failure mode i see is software developers, they click new chat, they 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, and then they say, make me an API, and all of a sudden they've got this backend API control with pink rest endpoints.

32:10You need to be thinking about a new chat, is the new array, each array... Chat hygiene. Chat hygiene. And before we go past that and just rewind a bit. There is proven scientific research of the notion of context rot and the existence of a dumb zone. So if you think that you've got like a 200k and like once you go like 60-70 % over that, like the LLM starts getting dumber. Actually dumber. So you want to minimize the amount of time in your dumb zone is what Dex Horty would say. And it is true. And so Ralph deliberately is you don't want to get in the dumb zone and you don't want to compaction, but you want to deterministically allocate.

33:00It's an orchestrator pattern for a gas town or for what is to come next, or someone can use to build their own gas town. AI has changed how we build software, but faster code doesn't always mean faster delivery. That's why Linear B is launching the Essentials Plan, your toolkit for measuring and improving AI productivity. You get the new Linear B MCP server, AI Insights dashboard, and developer surveys, all designed to reveal how AI tools impact delivery speed, code quality, and developer experience. Stop guessing which AI tools work and start leading with data. Visit LinearB.io to learn more and unlock the next chapter of AI productivity.

33:46you know let's let's talk for a minute about about gaston you bring up stevie ag and and maybe how gaston is uh something that's building on the idea of ralph maybe it's 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 you know 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 lightweight, and to be interchangeable.

34:19I think all of that plays into people building their own types of rafts, their own loops. And maybe Gastown is an example of one of those coming into fruition. But, you know, I personally, as someone who's building and using these tools too, I see it as inspiration to go and maybe craft my own Gastown. 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. Yeah, I used to work with Yegi over at Sourcegraph. I spent a bit of time traveling around APAC telling people, please pay attention, pay attention, pay attention, because I was sitting on Ralph.

34:59And Yegi was out doing much more worldwide big tours because he is Yegi. And he's just saying, And folks, pay attention. The orchestrator's coming. And he'd already done his rant saying this is going to happen. And I don't think people took him seriously. So on New Year's Eve, he dropped a blog post, said, Welcome to Gastown. Do not use Gastown. Gastown is not for you. But by the way, here's this beautiful chart that allows you to measure where you are on skill level. Mm-hmm. Mm-hmm. And in the figure five is the notion where you start deterministically allocating a agent and roughing. And when you get to like six and seven, six and seven, six is when you're starting to like learn to run multiple of these at the same time and then you experience failure domains when you're running two at the same time and then you realize you've got to do some engineering.