Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

Code and the Coding Coders who Code it

Code and the Coding Coders who Code it artwork

Episode 66 - Live (at the time) At Rails World 2026

Oct 6, 202645 minAudio

That RailsWorld keynote didn’t just start a debate, it forced a reset on how we talk about Ruby on Rails in the age of AI. We’re live from the Buzzsprout podcast booth with Joël Quenneville (principal developer at thoughtbot and Bike Shed co-host) and Tom Rossi (Buzzsprout co-founder) to unpack what’s changing when many of us barely write code by hand anymore and instead “coach” an LLM agent. We dig into the practical side of AI-driven development: why Rails conventions still matter, why generators suddenly feel new again, and how token efficiency and consistency become real engineering constraints. Then we get honest about software design. Do classic abstractions and object-oriented patterns help agents reason, or do they just get in the way when explicit code ships faster? We trade stories about broken integrations, duplicated methods that still pass tests, and why naming and domain language remain a steering mechanism for both humans and models. From there, we zoom out to the future of software products. If personal AI assistants become the primary interface, the “UI” might shift toward CLIs, MCP-style structured endpoints, and agent-friendly actions instead of screens and forms. That opens big questions about APIs reopening, who controls the agent layer, and whether we’re about to repeat the same platform power cycles we saw in the API boom years. If you’re building on Rails, experimenting with LLM coding tools, or wondering what your job looks like when the craft shifts from typing to verification, this conversation will help you think clearer. Subscribe, share this with a friend from your dev team, and leave a review with your hottest take on where AI is pushing Rails next. Send us some love. Judoscale Autoscaling that actually works. Take control of your cloud hosting. Honeybadger Honeybadger delivers best-in-class error tracking, intelligent logging, and Just Enough APM™. Disclaimer: This post contains affiliate links. If you make a purchase, I may receive a commission at no extra cost to you. Support the show

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show. Tap a time to play from there.

SPEAKER_00Hello everyone, welcome to Code and the Code Encoders Who Code It. We are live at Railsworld in the Buzz Sprout Podcast booth, and I am joined today by Joelle Quinville and Tom Rossi. Guys, do you want to do a quick introduction for those who don't know who you are?

SPEAKER_02Yeah, so I'm Joel Kenville. I'm a principal developer thought bot. I co-host the Bike Shed Podcast, and I'm giving a talk later today at RailsWorld on harness engineering.

SPEAKER_01My name's Tom Rossi. I'm one of the co-founders of BuzzSprout. Higher Pixels is actually the name of the company. Buzzsprout's one of our products for podcast hosting. Awesome.

SPEAKER_00It's what we use. So, guys, Rails World day two. Day one had a bit of an interesting start. What? There was a something controversial. I don't know. In shocking fashion, there was a DHH keynote that people won't stop talking about. So, Tom, I loved your reaction when I asked you how your Rails world was going, and you went, that fucking keynote. Okay. Whoa, wait a minute. Did I say that? That is what you said. And it's interesting because I'm hearing a lot of different opinions. Some people are like, well, yeah, he's 100% right.

SPEAKER_00And he may or may not be. It was just a very interesting way to start a Rails world where the tagline is the future of Ruby and Rails, with basically saying, Hey, yeah, I'm not using Ruby on Rails for much anymore. I only use AI. That's not exactly what he said, right?

SPEAKER_01No. I think that there's a little bit of overreaction on both sides, right? Because you've got David in David style, comes out really strong, right? And so then you've got people that are coming out really strong against it. But it doesn't make sense to build an SMTP server in Ruby. Can't we agree that it would make sense? But if I have to hire a developer to build it, man, it's worth it to pay the extra money and do it in Ruby, even though it's gonna not be as efficient because I don't want to hire somebody who knows how to build. But now with AI, it just changes the economics. So, and there's still a role for web applications and the idea that he's gonna switch over to Rust, right?

SPEAKER_01Like that's not really what he's saying.

SPEAKER_02No, it makes for a great controversial soundbite though. Yes.

SPEAKER_00I think that's a very DH way of doing it, right? He says something that people are going to talk about. He likes people talking about the things that he says. And I have heard a lot of other people defend that take in that, well, yeah, Ruby on Rails allowed them to get hay out quicker than if they had started with Rust and Native from the get-go. It might never have gotten through the app store either.

SPEAKER_01Like the you still have the app store rigor moreau. Yeah. The ability to be able to do that with turbo native is still a powerful argument for doing some of that stuff on the web.

SPEAKER_00Yeah. So the way that it's set a little bit of the tone has gotten the talks. I've spoken to a lot more people at this Rails world about how are they using AI in their development rather than what are you doing with Rails or what are you doing with Ruby. So given how the keynote landed and those conversations, what do you think is Ruby on Rails' role in AI development or in development in general now that we're all using AI going forward? Joel, what do you think?

SPEAKER_02I mean, definitely I think we can't minimize the role of AI. I have barely written a line of code by hand in the last six to eight months. So that is a transformative part of our industry, and that is absolutely something that we should be talking about. That's one of the big themes coming out of this conference. One of the things that I think is still sort of out there is how do we get AI to write good Ruby on Rails code? Code that maybe we're proud of, code that does the things that we want without a lot of the downsides.

SPEAKER_02And that is really interesting to see how that conversation has been evolving around the conference. I've heard a lot of people talk about generators, which are a sort of classic, almost original Rails feature that I feel have not had that much airtime in the DAS 10, 15 years. And all of a sudden people are reaching for them again because agents are able to build code that's much more consistent, more correct if you lean into that. And that's just one of the many techniques. And so bringing some of those things out of our minds, things that are tacit knowledge in our teams, and encoding that in a way that not only agents can use it, but then also all the other members of your team, I think is going to be a win for everybody, agents and humans.

SPEAKER_02And I think that's a theme that has come up in bits and pieces in a lot of the talks I went to yesterday. Uh and I think it came really strong in Robbie's closing keynote last night. And that's the thing that I'm going to be taking away from this conference to the work that I do in the next year.

SPEAKER_01Tom, what do you think? One of the most telling things during the keynote, which I think underscored exactly what he was saying, was when he asked the question, how many people have written code? What was it, like in the last week that you were actually writing the code? It wrote the line of code that you weren't telling an agent what to write. And there was not that many hands that went up. Like there's a lot of people. And for myself, I know it's been a while since I've written a single line. I've coached my agent in things that probably don't matter.

SPEAKER_01Like, I don't want that to be one line. I think that should be two lines. I think that should be a private method. I think you should rename this variable. And you know, it's like this guy, this guy thanks, genius. I'm so glad you're here to help me with your code, right? But it seems inevitable. The things that we're talking about, that AI is going to do more and more of the things that we do. And the question becomes what is the role of beautiful code? One of the slides that David put up, I'm really thinking hard about, which was it was something like abstractions, are they necessary, or something about abstractions?

SPEAKER_01And you know, you and I, we've talked a lot about object-oriented and saving maps. And I love, I love object-oriented thinking. I love object-oriented code. I love thinking through those abstractions, but it gets in the way of the AI agents. They actually do better when it's more explicit. Like a God object is a big no-no for us, but AI is like, hey, like I can read it all in one context window, right? It's all right here. And so, what does the future look like when beautiful code isn't what we think as beautiful code?

SPEAKER_01I had an issue. So, Bus Route, we do integrations with a bunch of different providers. We integrate Spotify, we integrate in terms of analytics. We get analytics from Spotify, we get analytics from Cloudflare, from CloudFront, from different CDN providers, right? So for me, I had written this abstraction to pull the data in because I'm a good object-oriented programmer, I think. And so I had written this abstraction, whatever. Well, then we did this integration with Spotify and it broke. It was just different enough that it didn't fit my abstraction. And I'm struggling with it, and it's right around the time when you know AI was everybody starting to use it. And so it's like, okay, I'm just gonna use it to AI, see what it comes up with.

SPEAKER_01You know what it did? It's like, why use an abstraction? Because it clearly doesn't work. So I'm just gonna write the code. Yeah, some of the methods got duplicated, but it didn't matter. The AI didn't care, and I'm like, that's not beautiful code, but I was done and it worked, and it was tested. And if it changed, the reality is I'm probably not gonna be the one to change it. I'm probably gonna have an agent do the change. So what happens? But like you were talking about, I can absolutely use an AI to write beautiful code, beautiful Ruby on Rails code, beautiful object-oriented code.

SPEAKER_01I can coach it on Sandy Met's thinking, right? But I just don't know how valuable that's gonna be.

SPEAKER_02So I think I might have the opposite take here. I was gonna say, I I I actually bring it up.

SPEAKER_00No, I've I have thoughts. I don't know if I have the opposite opinion, but yeah, I want to hear what you have.

SPEAKER_02I have seen, at least so far, that agents and humans seem to converge on a lot of the same best practices. The things that help me be more productive often seem to help an AI be more productive. Some abstractions maybe don't matter as much, but a lot of them still do. And agents are often better able to reason about your code if your abstractions sort of better describe the things that are happening. So some sort of low-level ideas from domain-driven design, I think, still really important because having the names, having the entities in the system allow the LLM to reason at sort of a higher level of abstraction.

SPEAKER_02I think the fancy term that people throw around for this is ontologies and being able to teach your agent that is huge. I recently worked on a project where we're processing a bunch of data and it's sort of going through this like multi-step thing, and it's not working well. A lot of things are not being processed correctly, and I threw an LLM at it and I was like, here, I'm giving you the real data. Look at the database, look at some documents, look at the code, and help me understand why it's not working. And it gave me a lot of contradictory answers and wasn't able to solve it for me.

SPEAKER_02What I did as a human is I came in and I brought in a mental model of what was going on. What we're working with is a pipeline. Things go one step at a time, and I've named these steps and what are the conditions at each step. And then it's a leaky pipeline. You can bail at various points of this pipeline and for various reasons. And I wrote a script that encoded all of this and could classify my data. And all of a sudden, I fed that to the AI and having that abstraction just skyrocketed its intelligence level. Same model, but all of a sudden now I'm getting much more useful outcomes out of it.

SPEAKER_02And I think that's in part because it has sort of stable definitions of what these populations are. It's also because I gave it language to think and reason in terms of pipelines and phases and bail points, instead of just like, I don't know, there's a column here, and it looks like this record doesn't have a corresponding Boolean in this table over there. That didn't mean anything.

SPEAKER_01Yeah. Let me be clear. I think abstractions are still gonna be around, and domain-driven language is gonna be more important than ever because we need to be able to communicate via English the requirements of the system, what we're trying to accomplish. So I think that those things are gonna exist, but I don't know, and I'm not saying I've like I said, I love object-oriented coding, I love what I get to do, but I can envision a day where it's like I don't need you anymore. Like I can just write the code and I'll write it in a way that is my own language. And for example, naming methods, like I'm always coaching it.

SPEAKER_01I'm uh don't call it this, call it this. But does it matter? Like, yes, certainly it does. The name of the domain concept matters, but the name of the private method that it's using to be able to do some calculation. If a human looks at it, I need a good name. If a LLM looks at it, does it really need it? Because it can actually read the method to know what it's doing.

SPEAKER_02I would say yes, because code is a steering mechanism for future runs of the model. The model's gonna read that code, and anything you write there is going to nudge it in a particular direction. And so I think our identifiers still matter.

SPEAKER_01So you would say you think that the role of the developer is gonna continue to be steering these models into writing well descriptive code.

SPEAKER_02I don't know if that's gonna be our primary role. We might have ways of fully automating this, but I think that sort of well-factored code, good abstractions, good naming currently still matter for models just as much as they do for humans. And right now, I think they kind of converge on the same patterns and principles. Now they may diverge in the next generation of models. I can't tell the future, but as of right now, it seems like we're still sort of converging. And a good heuristic is what's good for a human, is probably also best for a model.

SPEAKER_00I've noticed too when my agent tries to name things, if I haven't explicitly told it what I want it to name, or the domain concept is still new enough that the naming isn't well established. Sometimes the names it'll come up with tell me that it doesn't understand what we're building, it's naming something incorrectly. And not incorrectly as in, like, oh, that's the wrong term, but like the way that the class and method name flow together. It's like, okay, you think we're going in this direction. We're actually going in this direction. This is a good place to pause and make sure that you have the context.

SPEAKER_00And and going back to something you said earlier of like the God object, uh AI can hold that whole thing in its context. It's like that was the whole reason why we didn't want to do that in the beginning, is you would have to hold the whole object in your head, in your human head, right? But AI still has a context limit, it still can only hold, and we there's tons of research, and and I look into this a lot too about like as that context window grows and gets bigger, we get rot in its ability to reason. So keeping it tight, keeping small objects where it only has to focus on that small thing appears, at least in my little bit of experimentation.

SPEAKER_00Like the AI improves its reasoning about the code when it's got a small slice to deal with. And especially when we have like well-composed classes, it doesn't always open the modules unless the module has something to do with what it's looking into, makes a smaller context window, keeps the reasoning better. Yeah, again, we don't know what the future is gonna be. We're gonna get like 10 million token context windows, and none of this will matter.

SPEAKER_01Maybe probably even today, the context windows are so large. Like I think about the god objects in my systems, which I have some. That's not an issue. But I imagine if you give in the future, I don't know, it could be 10 years, it could be 10 weeks. I have no idea, but you give it the domain. This is the thing that I want you to do, right? At a higher level of abstraction, and you don't dictate the language. What's it gonna do? Would it reach for an object-oriented language or would it reach for assembly? Would it reach for C? Would it reach for something at a lower level?

SPEAKER_01You know, I don't know. But what I do know is that the way that I was interacting with AI a year ago was limiting its ability to help because I was telling it functionally to do something rather than giving it the context for why am I doing it? You're micromanaging it. I'm micromanaging it to the point where I'm getting methods that are named wrong because it doesn't really know. Like, dude, what are you trying to do? Why don't you just tell me what you're trying to do and I'll figure it out? But I was more like, hey, I need a method to do this calculation, da-da-da. But it doesn't know why I'm doing it. Well, now I'm much more likely to say, hey, let's build a plan, let's talk about from the user's perspective what the end goal I want to be able to achieve, and then backing into the actual code and it changes the way that the code is written.

SPEAKER_01You know, I get much better code, but it's just a different way of interacting with the AI. Yeah, that's interesting.

SPEAKER_02What I've liked doing is sort of like you're saying, giving it the goal up front, but then I've also wired things into my harness to create some form of back pressure so that there's sort of self-steering that happens. So that might be pulling in various code quality metrics, it might be running something like Rublecop or a test suite, it might even be something more agentic where it's applying some sort of heuristics, and all of those help sort of guide the outcome towards something that I would like better.

SPEAKER_01Your approach may 100% be right because what you're doing is you're getting the AI to write good human code. Yes. And that might absolutely be right. I'm saying, I don't know, man. I am so unsettled right by where we are as an industry. I'm so unsettled by what we're seeing. I'm excited about it as a builder, as a maker, as a business owner, as somebody who wants to deliver value to my customers and have them compensate me based on the value that I can bring them. I'm so excited about the future of AI. As a lover of coding, I am very unsettled.

SPEAKER_00They talked a little bit about that in the fireside chat this morning of like anybody who was identifying themselves as a maker or like a software writer or just someone who was just building and the mechanism they were using to build happened to be code, is probably pretty excited about AI building. Anyone who was getting more of their identity from the code side of things, maybe they were saying I'm a software engineer or I'm a software developer or I'm a coder, is probably having a lot more trouble with this because the identity portion of I write code is gone.

SPEAKER_00Am I still a Rubyist if I don't write Ruby? There's a bit of like going through the stages of grief with it. Like, I loved writing code, I still enjoy it. Like, I still uh wrote some code the other day and it was fun and it was like cathartic, and like I enjoyed getting to sit down and like fuck AI, I'm not gonna use it. But like, yeah, but yeah, my job is. They still ride horses and they love it. Sure, sure. You know, my job is to, yeah, you know, riding horses is great. But if your job is to get from point A to point B, you're taking a car or a plane or a train, not a horse.

SPEAKER_01Something that he did I thought was really, really good was he showed the clip from the how to build a blog that many of us that's how I got into Ruby on Rails, was watching that. And what he did was he picked up on the line, look at all the things I'm not doing. Right. Yeah. And that's what really was attractive about Rails back in the day was look at all the things I'm not doing because I remembered all the stuff we were doing in.NET, like just all this unnecessary code that wasn't, it didn't do anything. And so that kind of framework of thinking is Rails.

SPEAKER_01And so you can apply that same framework of thinking to the agentic world that we're moving into, which is look at all the things I'm not doing. I'm telling it business problems and I'm having it solve them, I'm having it do these things that normally I would have to do. I would have to go chisel, I'd have to go chisel the code, look at all the things I'm not doing. I'm not doing that anymore.

SPEAKER_00I'm just telling it go write a unit test or whatever. Yeah, and it was interesting, always trying to guess what he's gonna talk about and DHH is gonna talk about in a keynote, opening keynote for Rails World. How is he going to link up that we are doing AI-driven development and Rails? And I think that was the weirdest part about the keynote, is like he was kind of just like, oh, well, the Rails part doesn't matter anymore. We can write it in whatever, we can do whatever. Which is like, yes, it's true, but I think that even the AI is building faster when it uses Rails. It's got all of the conventions spelled out for you.

SPEAKER_01He showed how the metrics that are being put together by the evil Martians to show how good Rails is is a framework for AIs to build, especially when it's building code that we want humans to be able to read. That Ruby on Rails is still an incredible platform for building web applications. I just was surprised it wasn't more of the focus.

SPEAKER_02It's got some of the abstractions that we already want in there. I was talking about generators earlier. That's the thing that's built into the framework from day one. And those are great because not only they allow you to encode some of your conventions to get better code, they're also way more token efficient because the LLM doesn't have to generate all that code itself. And it's going to be way more consistent, and it's probably going to be faster. All properties that we really care about as engineers building in a system. And so, yes, you could work in another language and maybe would build its own little generator system, or it would just try to make up the code on its own.

SPEAKER_02But we get that for free in Rails from the starting line. Yeah. So that's the thing that's really nice.

SPEAKER_00Uh I want to shift gears just a little bit. I've been thinking and not talk about AI. Uh no, sorry. I want to shift gears to a different portion of thinking about building with AI, and that's less of how are we building software with AI and more of the software we are building with AI. I've been thinking a lot about that. We're potentially not building the right software anymore. I mean, some of the software we're building will always exist and we'll always need, but like what's an example of building software that we don't need? Okay, so a lot of CRUD apps is a form, right?

SPEAKER_00So structured information that gets sent to the back end, and that backend sanitizes the information and does whatever transformation it needs before it's saved in a structured database. But with LLMs, we don't need that level of structure anymore. An LLM could take a paragraph saying, Hi, my name is Drew. I was born on this date, and the like you don't need the structured form anymore, you barely need the structured backend anymore because the LLM can just go, Yep, I've got this and I've extracted what I need. Maybe you need the database because that's I'd argue that SQL is still a better way than NoSQL.

SPEAKER_00But I'm wondering if we're still thinking it's too structured, like even the AI agent, when you think of it as a program, like we define its tools, but we don't tell it what tool to use. We let it see what tools it has available, understand the task at hand, and then pick a tool, run it based on the results of that tool, do another run based on the tool, the results of a test. Oh, I'm not done, I need to go. And it changed from we're defining the workflow to just we're defining the tools, and then the LLM builds the workflow on the fly.

SPEAKER_00And I feel like software isn't that yet. We're still building the workflow in software. I'm just wondering if you either of you had a thought on potentially what the next version of software looks like. Not that we're building with AI, I think we're gonna be building it with AI. But what thoughts do you have about the future of the types of software we're building?

SPEAKER_02Are you suggesting that the days of the user interface are numbered, at least classically how we think of it, and that everybody will be interacting with a text box or maybe even a voice prompt?

SPEAKER_00Yeah, potentially, or even that the days of building a static UI are limited, and that the person will be able to describe the UI they want for whatever software they're using. Because why do you need the button has to be here? Well, I'm left-handed, I want the button on the left. You don't need to toggle that setting. You just say, Hey, that button should be on the left, or I'm left-handed, and like little things like that, all the way up to like I'm journaling, I'm going on a hike. I'm gonna journal, I'm gonna put in this box the longitude, I'm gonna put in this box the latitude of my start point, and then that for the end point.

SPEAKER_00And I'm gonna name the trail and I'm gonna link it to this, and then I'm gonna write this little description. I'm gonna fill in this box what the weather was like. I don't need that, I can just write down, oh yeah, I went on this trail, I started here, but like in natural language, we're coding in natural language now, but our apps aren't as natural language yet.

SPEAKER_01I like this way of thinking because we're trying to blow up everything that we know about how we interact with software today, and it may very well come to pass that everything changes. And I don't know if you guys keep up on what's happening in AI, but uh like I feel like Grokbot was a pivotal moment in AI. Do you remember when the iPhone came out and you know, all the soccer moms started buying it? And then all the BlackBerry users were like, Well, we've had this forever. Well, yeah, but a soccer mom wasn't using a Blackberry.

SPEAKER_01And once the soccer moms adopted it, it changed the world. It changed the way they interact with your phone. And Grokbot to me is a shot across the bow of what a soccer mom AI looks like. Muse came out shortly after. So I'm working with both of them, and it's starting to blow my brain exactly the way that you're describing, because I'm saying things like hey, I need to ship my skateboard from Jacksonville to Austin. Can you set that up? I don't ever want to see your interface. I don't want to log into your form. You know my address. You know my name. I just want a secure way to pay.

SPEAKER_01But what does the future of software look like? If these personal AI agents, if that comes to pass, soccer moms are no longer interfacing with a web browser and instead they're telling, hey, I want you to order this package from Walmart or whatever. I think there is a chance that what David said at the end, if you don't have a CLI for your product, you know, have it by the end of next week. You know, he's kind of throwing it out there. But there is a truth to if you have a CLI for your product, then any LLM can interact with your product without the need of an interface.

SPEAKER_01And there's no user going to that CLI, it's an LLM that's using it. And there is a real chance that that happens. I was thinking about for BuzzProut, we AI videcoated a CLI and we were debating whether or not we should even use it. And the more that we talked about it, like absolutely, because now you could just tell your agent, upload this episode. And then you could say, hey, how are my downloads looking? Which episode is doing the best? What should I do more of? What should I do less of? And now you're allowing me that intelligence to all happen within your agent of choice interacting with my software, which is really a repository in a CLI for your agent.

SPEAKER_02I'm curious if either of you are heard of the web MCP protocol. The general idea is sort of turning your website into a sort of very lightweight MCP so that an agent can interact with it structurally rather than trying to like scrape it or interact with it through computer use. And that's a little bit sort of, I guess, an alternative to making custom CLI that people have to download. They can just point their agent at your website and then all of a sudden they know here are the five actions I can take, and the agent can then interact with your website and do the things it needs to.

SPEAKER_01Ooh, that's interesting.

SPEAKER_00I would much rather do that than uh CLI. It almost sounds like what we would do for screen readers for accessibility. Yeah. Like this thing is decoration only, so the AI shouldn't bother look like it's almost like we just need the AI to be able to use the accessibility features of our sites instead of trying to actually like drag the mouse to this button or whatever. Like, yeah, here are the links you got, and here's the button, and here's the description of this image in case it matters or whatnot.

SPEAKER_01That could very well be a way for the future to interact with your application where the UI isn't on you at all because everyone wants their own UI. I just recently I was messing around with Fidelity Investments, their interface is just absolute garbage. Yeah, yeah, right, exactly. And so you know what? I was like, just gonna build my own. So I was like, these are the things that are important to me. I tell an AI, these are the things that I want to see. And it built a whole interface and reverse-engineered their API.

SPEAKER_01They don't have a public API, but they have an app, right? And so it looked at the traffic from the app to figure out how to make its own calls, and then it found other ones just by guessing based on the restful conventions that the the application was using. And so next thing you know, I have my own Fidelity app, right? That's totally for me. I mean, I do trades with it, it manages, it does everything in my own interface.

SPEAKER_03Yeah.

SPEAKER_02I've heard people tout Vanguard's UI being so bad it's good because it prevents customers from doing dumb things like trading away all their savings. Vanguard's sort of whole value proposition is sort of the not quite rest invest, but you put it in an index fund and you let it sit there and you don't touch it.

SPEAKER_03Okay.

SPEAKER_02And then people want to play around and trade. And typically, most people are not going to beat the market. And so I've heard people sort of tout that as being it's so bad and it's good because it prevents you from actually trying to trade.

SPEAKER_01Yeah. And I prefer to control my own destiny, which by the way, I've lost money. So this is not financial advice.

SPEAKER_00Please contact a financial planner for all your financial. Yeah, I mean, it look like you said, you built it, you built your own UI.

SPEAKER_03Yeah.

SPEAKER_00And I mean, to a degree, you're a tinkerer, you're already a software builder. That makes sense. But even when like the ClaudeBot came out, which I guess is like the original version of the rock bot and muse, of like Blackberry.

SPEAKER_01Yeah, like you're a coder, you're familiar with it.

SPEAKER_00But if you're a soccer mom, yeah, you were never setting that up, you weren't connecting your Telegram account to it or what have you.

SPEAKER_01But you probably have a meta account, either Facebook or Instagram, of course. And so Muse is gonna just start to move in that direction of now soccer moms are interacting with LLMs the way that developers are.

SPEAKER_00Yeah, and some of the things that you were saying, like, oh, I need to ship this to this. You back in the day, and back in the day, yesterday, you might pull up the UPS app or the FedEx app to do that, and now you're like, No, I just I need to do this, so that's where I'm like, I wonder what the next versions of software look like. Some software will need to exist because the AI will need to interact with it to accomplish tasks.

SPEAKER_02All the back ends still need to be there, right? You're you're talking about the front ends, the part that a human would typically interact with via UI.

SPEAKER_01Yeah, it's like we need a new term instead of UI, it's a AI U. No, AI. We need an AI interface, right? Whether it's the MCP, whether it's CLI, but instead of building a user interface for a human to interact with, let's be honest, I think there's a good chance that in the future it's gonna be agents doing all the work on the internet and you're just prompting it to go do whatever you need to do.

SPEAKER_00Right, which definitely changes the way we interact with the internet. Yeah.

SPEAKER_02What's interesting right now is that we're sort of in this transitionary period, right? We think it's probably going to find an equilibrium point at this place. We don't know for sure. And right now we're in that sort of transition, and so we need both. I can't drop a UI from my product today. That would be terrible for all of my users. So I need a really good UI, but I also probably want to have the CLI or the Web MCP or whatever it is, the next thing that we're developing, so that those sort of early adopters can use that. And then if we do sort of swing towards that new future, we're at the vanguard of that.

SPEAKER_03Yeah.

SPEAKER_02Then I feel like that's kind of the same for the conversation we had earlier about does the code need to be good for a human reviewer or someone who's reading the code? There might be a future where I never look at the code at all and I it can be total gibberish. Right now I'm kind of half and half. So, yes, I'm optimizing for code that's good for the LLM to write and to understand for itself, but I kind of still need it to be okay so that when I drop into it, I can follow what's going along. And so where I'm in still that like intermediate mode where I'm playing both sides. Yeah. And I feel like the same is happening in UI.

SPEAKER_01It's an inflection point like no other. The closest one I can think of is the mid-90s when the internet was first starting to impact businesses. And we could have never anticipated. We thought it was crazy how things were changing then, but it was slow compared to what's going on right now with AI. Yeah, right. I mean, I remember in the the mid-90s when Amazon launched, and bookstores are like, what is that? Who's gonna buy a book off the internet? Like, bookstores are going nowhere. But nobody could have imagined that Amazon would be selling us data storage and compute.

SPEAKER_01Wait, the bookstore is gonna be selling, right? Imagine if you could go back and talk to mid-90s version of yourself, right? Like, like you would have no, you could not paint a picture that they could be able to comprehend. Oh, yeah, yeah. The bookstore, you're gonna start using them for all of your tech stuff. And oh, yeah, by the way, you're also gonna order groceries from Amazon.

SPEAKER_00Yeah, it's definitely weird. I hope that's something that AI actually helps with is like the idea that like everything is Amazon or everything is these large companies, because I think that the mom and pop companies end up doing things better. Amazon just does it at a scale in a way that touches more people, reaches more people, but like in reality, like interacting with a mom and pop shop so much better.

SPEAKER_01The hypothesis that I threw out is that I think personal agents are gonna somehow enter the market in a way that changes the way that we interact with the web, which opens the door for hey, I need to order Doritos. I'm not necessarily gonna tell the LLM where to get it. Right. I could say what's the best price.

SPEAKER_02You're not going to micromanage your LLM.

SPEAKER_01I'm not hopefully you're not, but if you go, hey, order Doritos from Amazon, now you're micromanaging it, right? Um, and so I I just did this with I was playing around with Grockbach and I was booking my plane tickets, and I was like, I need to get from Jacksonville to whatever, look at my calendar. I was going to a conference, to a podcast conference. I said, I need a ticket. It looked at my calendar, it went to Google Flights and it was looking there. And then I said, Hey, make sure you're checking other places, you know, besides Google Flights, and you know what went and you know, try these different things. I was like, Oh, thanks, genius, right? Like use the same options, but I was interacting with it in a way that was just different.

SPEAKER_01And I was trying not to be prescriptive. That's why I told it, look at my calendar, and it figured out, oh, you need to be back Friday because your daughter's homecoming court is Friday night. Like it figured all this out where if it was me, it would have been hard to navigate all that.

SPEAKER_02I feel like I'm having a little bit of deja vu from the early 2010s. Remember when APIs were all the rage and everybody was opening up APIs and it, you know, you weren't going to be beholden to the UIs anymore because you were gonna have custom software and it was gonna talk to all the APIs, and we had like a golden age for a few years, and then all the companies started like closing down their APIs, and we kind of got to the world that we're right now. And now with the excitement around agents, we're sort of reopening up the APIs again because we want everybody to use their personal agent. Is it possible that we get into that cycle again where five, 10 years from now we're shutting down the APIs again, and everyone's trying to funnel us through their special agent or something rather than through our general purpose one?

SPEAKER_01Are we just gonna see that cycle again? It's gonna have to be some form of competition. Let's say that the personal agent, let's just pretend that Google becomes the soccer mom agent of choice.

SPEAKER_01Well, then Google, when you say, Hey, I want to buy this thing, Google owns the LLM. It can coach you to buy it from people that are paying Google. Right. You see what I mean? And so then all of a sudden, the open and closed conversation starts coming up of, well, now maybe I want to close it because I don't want this thing to happen. I can't envision it. It's scary when you think about those large companies having that kind of power, which they already do. I mean, when was the last time you actually did a Google search? I'm a boomer.

SPEAKER_00I I still do my Google search.

SPEAKER_01Yeah, I don't you don't need the AI response up there.

SPEAKER_00I try not to. I legitimately try not to, just because I think that if I get used to doing that, I will stop using my brain. Like I'm legitimately concerned that I'm just gonna stop my ability to reason about things and like consume information and form opinions, and I'm just going to be going, Oh, well, Gemini says this is how I should think.

SPEAKER_01It's gonna be it's gonna be amazing. It's gonna be wally. We're just gonna have slurpees and ride around on a cruise ship. Like just embrace it, dude.

SPEAKER_00Maybe get an elastic ban waste. Yeah, no, I mean that's the thing. Maybe being resistant to the change is is the wrong way of doing it, but that's why I'm so interested in like what is the next iteration of software look like? How are we going to be building? Just because I'm hoping to be able to prep for it a little bit.

SPEAKER_02One thing that came up two days ago with the Rails at Scale Summit. I don't know if either of you have showed up for that, was how to think about working with LLMs as developers when we're forced to use smaller models. Because earlier in the year everyone was token maxing and it was just sort of unlimited budgets for everything. And now there's a little bit of a spring back where people are being asked to maybe consider their budgets and to spend a little bit less. And so some of the questions that are coming up are like if we're doing this at scale, where all these developers are running multiple agents and doing all these tasks, maybe we want to use some of the good but smaller models, like Luna.

SPEAKER_02And how do we get good value out of those? What sort of tasks are they good at? How do you scope things so that they don't sort of fall into their bad patterns? Have either of you experimented with working with perhaps state-of-the-art models, but that are not like the top flagship model?

SPEAKER_00A little bit. I flip around models depending on tasks. I also have been tinkering with running a local model on my desktop computer. Okay. Which is slow, but I like the idea of it's like it's there. I'm not shooting an API request out. Right. So not to the level that you're talking about, but I am interested in using like almost the right LLM for the task so that I'm not burning my complete budget on one thing. How about you?

SPEAKER_01My hypothesis is it's gonna matter less and less. It's gonna matter less and less. Right now, I know which models I'm reaching for for certain tasks, but I use cursor as well. So I've got cursor, I've got chat GPT, I've got Claude. Cursor's auto function, I think is gonna be the future, right? Like, you know who's really good at picking models? Models, you know, like LLMs, open router, right? Like being able to choose. And so I think that that's it's gonna become less less and less important for us to manage it. I think that that's gonna happen.

SPEAKER_01And I also think that the costs will come down as long as there's competitors. What we need to make sure is that not one company owns the models, right?

SPEAKER_02And the open models, they're putting so much pressure on the industry in a good way for us as consumers.

SPEAKER_01Yeah, it's a good, it's good for us as consumers. And what we don't want is for one person to own it, because if one person owns it, then those costs are never gonna come down. But if there's competition, they're gonna come up with better ways. We've already seen it with the context window size always growing, the improvements that we've seen. And so I think we're gonna see those costs come down as they compete, but we're just not seeing it yet.

SPEAKER_02Have either of you experimented with Jed, the new sort of kind of classifier type LLM that's come out.

SPEAKER_00No, I was lucky enough to been on sabbatical at work, and I feel like I missed so much because I got back and all of a sudden there was a bunch of new models, and then everyone was talking about Jev, and I'm like, what's Jev? And they're like, Oh, it's this different, and like it was one of those my eyes glazed over, I just zoomed out. I was like, All right, does it help me build software faster? Well, no, it can help you do oh okay then right now I only have so much brain power to catch back up, so I haven't deep dove into Jev yet.

SPEAKER_01I've seen a lot, obviously, of conversation about it, but I haven't played with it yet. But my understanding is it's moving us in a direction of models that do specific things really well, and I think it goes back to the idea of the harness doing the work of picking right. Hey, based on the task that you're giving me, I'm gonna use Jev to organize the data, categorize it, and then I'm gonna feed that into this other model and it's gonna do this. Versus right now, I feel like I'm like copy, paste, copy, paste, you know, like moving it around.

SPEAKER_02Because what it's really good at is routing, and we can use it to sort of classify tasks into different buckets and then route towards a different path based off of that. One of those ways could be analyzing a prompt and then telling you what is the best model to route it to.

SPEAKER_01Yeah, let's see, that's pretty cool. So then imagine you tell it, okay, I've got a subscription to Claude, I got a subscription to open the eye. You go figure out which one to use based on the task I give it. Because right now I've been playing around with herder, and so I've got different tabs and I forget which one is on which model. Oh crap, this one is on Fable, and I just put in the prompt that I should have put in on the one, you know, and so it'd be great to have something like that.

SPEAKER_00The last thing before we wrap up, thinking about using a different LLM, similar vein, but what I'm looking forward to, what I'm hoping happens is we get almost inexpensive reinforcement learning where we're not just providing, like right now, have a pretty robust memory system using an obsidian vault so that my LLM can like look through things that we've run into and it has tagging and all that stuff. It's like, oh, we're working on this, let's look up tags and drill down an obsidian. But like actually fundamentally adjusting how the LLM is trained and all of its default information that it comes out of the box with, I think would be really a cool next step.

SPEAKER_00Like local LLMs that you have trained on how you want your software written, on how you want your code written, on how you prefer tasks to be done. Do you guys have any thoughts about that stuff or have thought about it or have a I hope this happens in the future?

SPEAKER_02In some ways, that future is already here. It's just a little bit more clunky, right? It's all the different ways that we can steer and it's all done via the harness. You can do things with commercial harnesses already. If you really like to tinker with this stuff, you can use harness like Pi, yeah, which I like to describe as like the Vim harnesses.

SPEAKER_01Uh you're really selling it there to the right person. Yes, yes, he is.

SPEAKER_02Yes. And so you can really tinker around what the model does because a lot of the things that we interact with, and we're like, oh, I want this behavior, I want this behavior. It's not really the model itself, it's clawed code or cursor or all the stuff around it. Even things like what is the size of your context window? Like that's more or less set by your harness. And like, how does compaction work and all that? That's not the model, that's your harness. So if you're willing to tinker, I think that future is already here today. It's maybe not exactly what you're saying, the like modifying the weights, but I think the outcomes are often in the harness rather than the model.

SPEAKER_01I'm finding myself wanting to be less and less prescriptive. So, like, I've gone and I've deleted out tons of those documents that I've created in the past because I want to see what it comes up with.

SPEAKER_02It's a great practice to do every now and then. Just wipe it clean, see what the current generation of models do. Where do I still need to steer?

SPEAKER_01Yeah, it's still so eager to please. I told you I built the Fidelity Thing Majaker. So I was having it do some trading because of course I would. And it was recommending and it recommended meta as a stock. And I was like, I hate meta. I don't want to buy meta. And so every interaction after that, it would be like, okay, well, I'd recommend this stock and it's not meta. And it'd be like, oh, I'm checking out all the stocks of what the current upsides are, but I'm not checking meta. And it was like, dude, calm down. Like, like, you don't need to tell me about meta every time. Okay, got it.

SPEAKER_01I'm writing a rule. I don't have to tell you about meta every time. And I'm like, oh my gosh, forget it.

SPEAKER_02I'm suggesting you a stock, and I'm not talking about you know what.

unknownYeah.

SPEAKER_00The stock that shall not be named.

SPEAKER_02Yes, yes, exactly.

SPEAKER_01And so I find that the LLMs are so eager to please. I'm nervous that I'm coaching it in a way where it's learning my bad habits. Um, you know, where I've definitely had it. I I just had this happen last week. I was fixing a bug and it broke the tasks. And I was looking at the test and I was like, why are you testing all this stuff? I'm like, you don't need to test all this, just make the assertions that you need to make.

SPEAKER_03Yeah.

SPEAKER_01And it was like, okay, cool. But then I expanded it on the PR and I saw why it did it. It was because it saw my tests above it. And those tests were written like 10 years ago and I was a moron and I had all these assertions that I shouldn't have had. And so the LLM was eager to please, and it was like, hey, well, this is the way you did it, so I'm just gonna do it that way. And it was nice enough not to point that out. But when it created its test, it just copied the same assertions, and I was like, ooh, I wish I could tell you. Assume I'm a moron. Like assume all the code that you're looking at, I'm a moron.

SPEAKER_01Do it the best way, not the way I've done it in the past.

SPEAKER_00Interesting. All right. Well, thanks, guys. Thanks for coming on to talk about a little bit about Rails world and the state of AI coding and Rails' place in it. Well, I'm looking forward to your talk later. It is on my agenda to make it out to. And I hope to have you guys on again to talk more about how we're building the next generation of software with AI because God knows we're not writing code anymore.

SPEAKER_02Does that mean you're renaming the podcast?

SPEAKER_00I'm just I don't know what I'm gonna name the podcast. It's gonna have to because we're not coding coders anymore. We're AI AI AIs, yeah.

SPEAKER_02AI agents, the AI?

SPEAKER_00Listen, I only get so many good names out at a time, and I feel like I wasted a lot of my naming budget on naming this show. Prompting prompt for the prompts that they prompt with. Something like that. No idea.

SPEAKER_02Hey Claude, rename my podcast.

SPEAKER_00I'm gonna I'm gonna have to just start interviewing AI. Like I'm just gonna have to have a podcast where it's me interviewing an LLM because they're the code encoders at this point.

SPEAKER_02Honestly, that would be probably a really funny joke episode.

SPEAKER_00Yeah.

SPEAKER_02You interview Claude, you know, one of the open AI models and Grok, and they're they're all having a dumb conversation with you.

SPEAKER_00One of my coworkers named his GPT Gerald, and he always will be like, Well, Gerald thinks this, and it's just the funny way that he talks about it. It does sound like he's talking to a human. I'm like, I wonder if I could interview Gerald on the show. That's good. All right, well, thanks, guys. This was awesome. And listeners, if you're not at Railsworld, you're missing out, and I hope to see you all soon. Bye.

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Episode 65 - Mike Coutermarsh

    Apr 21, 20261 hr 12 minAudio

    Your database is slow, your Sentry is screaming, and the backlog is full of “we’ll fix it later.” What if an AI agent handled the boring but high-impact work while you slept and just opened a clean pull request for...

    Credit the people on this episode

  2. Episode 64 - Delaney Gillilan

    Apr 7, 20261 hr 33 minAudio

    React made a lot of us feel powerful, then maintenance made a lot of us feel tired. Drew Bragg sits down with Delaney Gillilan, creator of Datastar, to argue for a different kind of “modern web” one that keeps the...

    Credit the people on this episode

  3. Episode 63 - Travis Dockter

    Mar 24, 202658 minAudio

    What if the most useful software in your life isn’t a product, but something you built for yourself in an evening? That’s the spark for this conversation with Travis Dockter, a Rails developer and organizer of Blast Off...

    Credit the people on this episode

  4. Episode 62 - Cameron Dutro

    Feb 3, 20261 hr 16 minAudio

    What if you could keep Rails pages fast, accessible, and SEO‑friendly, yet still get modern interactivity without shipping a mountain of JavaScript? We sit down with Cameron Dutro to unpack Live Component, a...

    Credit the people on this episode

  5. Episode 61 - Ernesto Tagwerker

    Jan 27, 202659 minAudio

    Rails upgrades don’t have to feel like crossing a minefield. We sit down with Ernesto, founder and CTO of FastRuby and Ombu Labs, to unpack a pragmatic path from legacy Rails to Rails 8.1 and how AI can accelerate the...

    Credit the people on this episode