Scott Hanselman on AI-assisted programming

Speaker A: And then do you have a hard stop time apart from what we saw?

Speaker B: No, I'm chilling. Uh, after this, I'm going to go on Facebook Marketplace and buy a Mac Mini. So that's what you're. Between me and my Mac Mini should be sitting on their porch.

Speaker A: All right. Yeah, I guess we'll get started.

Speaker B: All right, let's do it, man.

Speaker A: Today I'm talking to Scott Hanselman. He's the vice president of developer community at Microsoft. I last spoke to Scott back in 2021 where we discussed Net.

Speaker B: Hey, man, how are you?

Speaker A: Pretty good. Today we're going to talk about all these AI assisted development tools that are out there. And I think it's super overwhelming, I think for a lot of our listeners and for myself. So let's kind of give an overview of maybe the types of tools you have experience with. I think maybe we could start with just the chatbots. The chatgpts and the text box go back and forth on.

Speaker B: I would even go farther back, if I may. Yeah, uh, and like, not to put too fine a point on it, but like, let's go back to like 1975, because I was thinking about this yesterday. I was trying to explain this to my wife and it's like when I started coding in like 80, 84, I went from assembler, uh, to C and people would be like, c, come on, that's too high level, man. You got to get down to the metal, right? You're like, it's gonna rot your brain. C is gonna rot your brain. And then in the early 90s, we got color. And by that I mean we got syntax highlighting. And they're like, oh, it's going to rot your brain. And then we got stack overflow. And they were like, oh, you're going to rot your brain. And now we got this. When we started thinking about AI augmented coding, it was next token prediction. You type A for loop, you hit space, and then it makes like ghost text and suggests. So that's the first thing. Then there was the chatbot. So just right before the chatbot, there was the ghost. Ghost text. Uh, hey, you know, uh, here's, you know, it would like interrupt you and um, they would have like next token prediction within the editor. And some people really like that because it's like, oh, I know what I'm doing. And then it's just going to autocomplete. So it's like autocomplete on steroids. Then people went into chat GPT and started asking questions and then they would copy paste not From Stack Overflow, they'd copy paste from ChatGPT. Then the agentic loop started and that's where things got interesting. And if you really want to sound smart in meetings, uh, just say agentic as many times as possible. And agentic loop sounds even smarter. But the idea was that why am I copy pasting this code from ChatGPT when I could just let the agent the LLM run, build, look at the text, see a warning, fix the warning and then loop. And that's when we started getting into things like GitHub, Copilot, CLI and Gemini and AMP and OpenCode and Claude and all of those kind of things where it's LLM sees some tokens, gets the result, makes a judgment. And I call it the ambiguity loop. Because this is what loops. This is the thing about programming versus LLM programming. Programming is not ambiguous. It runs exactly as you wrote it. And if there's a bug, it's your fault. And if you want a for loop, you write the for loop. You do it in bash or powershell or whatever and it does the thing. But an ambiguity loop is like, yeah, I don't really know, like things could happen and we fool ourselves into thinking like regular expressions and stuff. Kind of like handle ambiguity. But if you imagine this kind of slider bar of ambiguity between, uh, I'm parsing this binary and it's going to work this way, and here's the way that the bytes are packed versus the like, I don't know, man, it's unstructured data. What are you going to do? LLMs are really good at ambiguity. Turns out the toil part of programming is really ambiguous. Like I'm writing this bot on this other machine over here. I'm running what's called a RALPH loop that we'll talk about later. And it's bumping into all kinds of little weird problems. And those little weird problems are just crap that I would have to go and Google or look on Stack overflow. It's just ambiguity. So I'm running not a RALPH loop, but an ambiguity loop. That's like working that out. Because that's not the fun part of programming. The fun part of programming is like building stuff. So I think that these agents and why people are freaking out about them is because it's that same feeling that we got when Stack overflow happened. We're like, oh, programming is over. Or when, when I got syntax highlighting and I was like, uh, I don't have to do anything now that all the text is in color. Does that make sense? That's how I contextualize it across a historical context.

Speaker A: Yeah, I think that's good to bring it back to the syntax highlighting and the autocomplete because I think nowadays it's very uncontroversial. People, they for the most part will accept like, oh, the syntax highlighting is there, it helps me.

Speaker B: Uh, that's a good way to point

Speaker A: out what I'm looking for. The autocomplete I can quickly see is the suggestion something I want, yes or no. I think for the most part people are on board and, and I think this ambiguity loop you're talking about is maybe a little bit. I think people have more mixed feelings about is because of that, what you're describing, where it's like this slot machine, Right? Like maybe it'll give you a good answer, maybe it won't.

Speaker B: Yes. And here's a really great. I had this internal conversation with this guy, like we're on teams and this is an internal guy. And he was so excited because he did a one shot. A one shot is when you type one prompt and then software pops out. He did a one shot and he cloned Minecraft. And you know, there's always this, there's this thing that, um, I think it's called like there's a paper. I think it's like Project Unicorn or something like that where like OpenAI cloned Minecraft. But forget about that result for a second and think about the simple semantics of typing a prompt. Make me a clone of Minecraft that runs in the browser using 3js. Which word in that sentence is doing the heavy lifting? The heavy semantic lifting. There's Minecraft and there's three Js because you're steering it and saying use three Js versus using Babel versus using whatever. But in that sentence, make me a clone means a copy of Minecraft means a lot. The word Minecraft semantically is effectively a 50 page spec. So I was trying to tell the guy, dude, that word is doing literally all the lifting. Try explaining to it to give you a clone of Minecraft. Except you don't get to use the word Minecraft. Right. And now that would be effectively impossible. Now you go from oh my gosh, I did a one shot of Minecraft to this is a hot mess. And I really need to be specific. And you're going to get, like you said, a roll of the dice. So we need to ask ourselves, when we are making these ambiguity loops, what is the job of the programmer? Well, it's to be specific because it's all about expressing your intent and making your intent understood. And I'm sure we would have loved in the 60s, 70s and 80s to like just talk to the thing, but it wasn't specific enough. So we needed to be so specific that I'm like, put that electron in that register and we're going to use assembly to do it. And now suddenly I can say, hey, computer, what's my blood sugar? And it knows to go into a REST API to get the tokens, to go up to my Mongo database to bring all that kind of stuff down, because suddenly the Lego pieces are high enough and the stack is deep enough that I can get what I want. And it might try it in Python one day, it might try it in JavaScript one day, but the result is I get to know what my blood sugar is because the REST API is called. But that ambiguity loop shouldn't be something that we should take for granted because it's not going to work every time. That's why I think it's interesting to use the ambiguity loop to get me specific software that runs absolutely reliably. Because talking to ChatGPT or talking to Claude, that's not a scripting language, it's effectively a prose compiler. And that's different. So I'm using it to make grounded scripts that I then run, rather than just yapping at the thing and pretending

Speaker A: that it's bash files in some ways. It's like that, uh, shift you were talking about earlier, where you went from writing assembly or machine code to writing compiled languages, compiled down to that. And then we got to higher level languages like JavaScript and then now we're one level. Even before that we have our English prose that's helping us compile or generate the high level language.

Speaker B: You know how people use like LUA or C and they build it into like Unity apps, right? And that's like, that's the scripting language for their game. That makes sense to me. To embed an LLM would be to embed ambiguity, which they think, oh yeah, look, the NPCs are going to be able to do stuff. But now they could do anything. They could say dumb stuff, they could go off the rails and the game could completely do something that you don't want it to do. That's what we don't want to do. We don't want to push out slop, because rolling the dice is not software engineering.

Speaker A: And I think maybe another important point is with that example where you said somebody tried to. One shot, let's Write Minecraft. Right. My guess is that the end result was not actually Minecraft. It was probably something, something that resembled it a little bit.

Speaker B: But yeah, I wish I could show you. It's actually, to be clear, it's super impressive. It's a full voxel world. It generates the voxels, it runs in the browser, the textures are all there. There's no crafting in mine. Um, but you know, you've got an inventory. It's. It's what you. It's Minecraft. If you woke up from a dream and you're just like, oh, I had a dream I was in Minecraft. So it's maybe like 60% of Minecraft and that's pretty cool. But again, the spec lives entirely in the word Minecraft. Because if you said, make me a clone of Fufu and Fufu isn't a thing, you're gonna just roll the dice and it's gonna go, it's gonna go insane. That's why it's not a reasonable demo to say, make me a Space Invaders, make me a Pac man, make me a Donkey Kong. Because those words do all the heavy lifting.

Speaker A: And I feel like some of the higher levels people who are running these businesses, CEOs and things like that, they have this vision that I'm going to be able to ask this rather ambiguous thing and get a complete product, but to me it feels like we're not actually there.

Speaker B: Well, so an example that I really like to use. Um, and may I ask, do you have a car and do you drive?

Speaker A: I do, yeah.

Speaker B: Okay. Do you drive stick or do you have automatic. Do you know how to drive stick?

Speaker A: I don't know it well enough to comfortably do it.

Speaker B: Okay. Do you feel that it is a personal failing that you don't know how to drive stick?

Speaker A: I do not.

Speaker B: Okay, so take everything that we just talked about and now apply it to programming. Right. Like, I could do assembler, but I couldn't do it without Googling anymore. Do I feel bad about that? No. Could I drive it in an emergency? Yes. Could you drive someone to the hospital and stick? Yes. You could probably make it happen if, if you were in a pinch. Here's the question, though. What if you answered just now, I don't have a car. I only know how to Uber. Then your relationship to the vehicle fundamentally changes. And if the vehicle breaks, you're going to get out of that car. You're going to leave that poor gentleman there, he's going to have to fix that car, he's going to have to repair it oil change and tires and you're just going to get in another Uber and fly away. But how deep is the stack for you to simply move your body from point A to point B and you have no control over how the guy's gonna get you there. So I think that there is value in driving stick occasionally, uh, trying to get somewhere without a gps because then the ambiguity has to be and the responsibility is on you, the honest is on you. While right now the idea like, hey, send a stranger in a car with candy to pick me up and I just like lean back. That's the same thing. Asking an Uber to send you somewhere and asking Claude Code or GitHub Copilot to make something for you with no specificity. You're lucky. I hope you get to the airport. Good luck. You know what I mean?

Speaker A: Well, I suppose in that example, the Uber or the Lyft, you're giving an address most likely, and it's sort of this well defined problem of I need to get this person to the airport or I need to get this person of the airport. And I feel like that's a little different than me saying telling an LLM, um, please build me a web browser or build me Minecraft.

Speaker B: That's fair. I think an example would be, uh, hey, Uber, take me to the airport.

Speaker A: Without specifying which one?

Speaker B: Without specifying and there's no ask user question.

Speaker A: So I guess your point maybe is that it's somewhat unrealistic to say something so vague and not expect there to be uh, a back and forth to be clarifying questions, that kind of thing.

Speaker B: Exactly, Exact. Exactly. And I'm sure you've been in an Uber and you've told them, hey, take the 405 because you are steering them. You're not driving the car, but you steered. And that's why we're finding steering in these CLI coding agents so delightful where it's like you're 90% in the direction I need you to go. Just get off on this exit. Those little moments matter. Specificity matters. I'll give you another example. My son is 18 and he has a depop. So he sells like vintage clothing on Depop, which is like a marketplace for people to sell clothing. But he ships on a website called Pirate Ship. He doesn't use the depop built in shipping. So I had this vision to make it so he didn't have to copy paste stuff back and forth. And I thought I was going to do it with like playwright and I was going to automate browsers And I was going to build a whole, like, dashboard for him to do shipping. And I was basically using the copilot as a rubber duck. I was talking back and forth to it. I was talking to myself in the mirror. I'm not actually talking to an entity, I'm talking to myself. And in the brainstorming, just as if I called you and brainstormed, we realized that, oh, I don't want to write a playwright extension. I want a browser extension. So I wrote a browser toolbar for him. And I wouldn't have come up with that. I would have gone all the way down playwright automation and been wrong because of my own biases if I hadn't brainstormed. And a little bit of the ambiguity loop popped up and said, have you thought about a browser extension? And it clicked. And then I went, immediately, it's like, hey, take me to jfk. You know, EWR has a direct flight. Have you thought about that? You're right. Let's do it. And that back and forth, as well as having a lot of experience and understanding the power of a browser extension turned out to be a really great experience. And now he's got one click shipping for his, uh, little business.

Speaker A: Very cool. I have people ask me sometimes, is it. Is it still worth learning how to code? Is it still worth learning all the specificity of HTTP or how the browser works or databases? And I think what you're maybe implying is that having at least some background in these things will help guide you when you're asking these questions, when you're trying to come to a solution. If you don't have that background, you might not get there.

Speaker B: I acknowledge my position in history, my age, where I grew up, I speak English. I'm, um, a white guy on the West Coast. Like, all those kind of things are the context in which I came up. I came up in a very interesting time in the 80s programming. So I don't want to sound like old man who shakes fist at Cloud. Hey, bro, why are you gatekeeping stuff? Right? But I see people saying, hey, man, I just vibed the great. The world's greatest new SaaS. Check it out. And they tweet localhost3000. And it's just like, uh, uh, right. So, like, do you, like, buy your furniture at ikea? I do. Could you cut a board? Did you take, like, probably wood shop? Yeah. You know, would you build your own furniture? No. But you have an appreciation for building IKEA stuff. I think that's why IKEA makes you build it. So that you don't feel like it just appeared in your house. It can't just like spawn up like Billy shelves. So yes, while recognizing that it might sound like, hey, this old guy wants me to write everything in assembler. Yes, yes, yes. Appreciating that oil changes are a thing, DevOps is a thing, build loops are a thing. Um, and quality matters. So I would argue that it's not the death of software engineering, it's the death of toil. And the only thing that matters now is taste and judgment.

Speaker A: So where do you think that balance is? What do you think people should be learning where they can use these tools effectively and they can still be building quality?

Speaker B: I think if I were. And by the way, when I retire, I'm going to go and teach, uh, high school computer science. That's my plan. I've got it all lined up. And I was a professor before and I'll be a professor again. I don't think there's enough computer science history and I don't think there's enough basics. I don't think anyone learned HTTP in college. They just kind of sat down, they did compilers, they learned Lex and yac, and then suddenly you're learning OSS and then you're writing apps, but you pick up HTTP and DNS and stuff like that. Maybe like I took a 400 level TCPIP class, but as a general rule you're not spending a lot of time learning how to do kubernetes, uh, in school. I would, I think HTTP DNS, nats, like that's all bread and butter. That's, that's sociology, right? Like why do you learn about world history? So you can go and be like a corporate drone. You learn because it makes you a whole full person. So yeah, there is someone right now Vi coding their bachelor's degree in software engineering and I love that for them. But they're going to get nailed. Just as people who don't learn about sociology and history and psychology are going to get nailed. The Dunning Kruger effect is a real thing.

Speaker A: Hypothetically, when you're teaching your computer science class, what would you tell your students in terms of using these AI assisted tools?

Speaker B: I would say it is a funny two faced thing because it is a senior engineer with infinite patience, but it is a junior engineer with infinite energy. The role that the LLM chooses depends on how you're coming at it. If I come at it as a net engineer and I'm competent in C, then it is a junior engineer with infinite energy and I need to Point it exactly and I tell it what I want and it does the thing. But if I come at it as a Python developer, of which I am a novice or a medium, then it's more senior than I am. So I would be less likely to disagree if it makes a recommendation. If it does something in. Net like, no, uh, don't use that, use this, I will say, no, we're going to jfk. Trust me, this is how we're getting there. But if I've never been to that country, I've never been to that state, I don't know which airport's the better. I'm going to take the advice of the Uber driver, but it's the same LLM, it's the same model. And if you come at it with that sense of like, no, I know what I'm doing, strong opinions weekly held, you're going to have a much better experience. Additionally, the opening prompts matter and I think that there's room for learning, uh, prompts. So if I were teaching the class, I would include a markdown file and skills that would force the user to like, oh, you just tried to vibe your entire final. Here's some multiple choice questions, here's some, you know, here's an exercise for the reader. I'm not going to let them just, you know, 1. Shotting your computer science file. Uh, you're doing yourself a disservice, I think, if you do that.

Speaker A: Yeah, that's an interesting point about how the way you interact with the LLM, um, is so different when you know the domain or you know the language well. And I wonder if, in your opinion, if it's, let's say you don't know it well, are you in a sense sort of faded to get a result that's maybe not as good, like, because you didn't have that background because you couldn't steer it.

Speaker B: If you don't understand why it did a thing, then I would not feel comfortable shipping it. And that's where I think the limit is. Like, if I were going to do something, like I'm doing a bunch of Python stuff and I think I may be a 4 out of 10 or 5 out of 10, but I know how their type system works versus TypeScript versus whatever. So I'm asking generic computer science questions like, you know, like, what's the Python equivalent of a linter? Do I have enough guards for my exceptions? Like, that's not Python questions, that's just quality, that's just taste, you know, and I'll say I'll challenge its assumptions. I built a blood sugar management system in Python a couple of days ago and I've asked it, should this be Typescript? I said, make me a pros and cons table about moving this to Typescript. I was like, no, I think this works. Python works great. And we kept it at Python. But then I had another thing that I was doing, um, in C and it was on. NET Framework. Um, I thought it would just be fine on. Net framework because I. No, no, we should move to net 10. Here's 10 reasons why. And it's a back and forth. I think too many people are trying to vibe code stuff in two or three shots while I'm doing stuff in 50 or 60 and having ongoing conversations. Like I'm using. I've got, um, I'm gesturing right here, people can't see to my terminal, which has 13 tabs open and they're all doing something right now. And I use Copilot Cli, which is like Claude code. It's an agent loop and it has infinite sessions. So it has this idea of like, the context of like how long it can think about stuff. Doesn't exist in Copilot Cli. So I can just go dash, dash, resume and pick up stuff I worked on a week ago. And it's like a hibernation. I just hibernated the agent. I brought it back. So it's remembering all this stuff and all the decisions I've been making on. And I'll say stuff like, you know, revisit our conversation from a week ago. Do you still think that's true? You know, go and look at this pull request, go and look at the stack overflow. Go and look at this new tweet that someone that Simon Willison said. And given that new context, what do you think? So I'm constantly challenging it. Uh, and it's almost like I've got 13 engineers that are either working full time or they're just asleep and I wake them up and I ask them a question and then I go off and I do my thing. And it's really been like, I feel like an engineering manager with a lot of really cool little junior engineers with so much energy.

Speaker A: And for someone who isn't super familiar with the domain or what the language they're working in, how do you verify the assertions that the LLM is making?

Speaker B: I love that. I love that because I'm not an expert in Python, but I wrote my blood sugar management system in Python. However, I am an expert in type 1 diabetes, so I thought to myself, how can I verify that this is valid? But I don't know the language. And it's, um, what I call Fear Based development. Right, it's fdd. Fear Driven Development. I don't want to screw this thing up. It's my blood sugar. And in the words of Brian Lyles, test all the time, right? Always. So how many tests do I need before I don't feel scared? Uh, in this one here, it ended up being 270. That's just a number. But like I needed about 70% code coverage across almost 300 tests. And it's on real data and it works so that that CI CD and it runs in GitHub Actions and it runs locally, it does nothing until it tests. And I would again argue that that's about software engineering, not about Python. So as long as there's a test framework and a loop, and that loop is verifiable, that's what's important. Because everyone's excited about RALPH Loops, which we should talk about at some point. RALPH Loops are great, but if you have a million monkeys and they're typing in a million typewriters and you're hoping that they're going to type Shakespeare, if you don't have a test that validates the Shakespeare, then you're just wasting monkeys and wasting typewriters.

Speaker A: These tests, are you generally writing them yourself or is this something that the LLM is also writing?

Speaker B: So I will write for four or five tests myself, then I'll tell the LLM to write me other tests on the boring stuff like edge cases. And then, because we know that in software engineering, uh, one of the big problems is naming stuff. What are the top, the top three issues in software engineering is naming stuff and off by one errors. So that is the things I watch for. I know that it's edge cases. So I'll write a test, I'll have it make me 10 more, and then I'll look for all the edge cases. And yes, it's true that I haven't deeply validated them. And it is true that depending on the model you use, it'll comment tests out, it'll hide stuff from you, it'll pretend, it'll gaslight you into thinking the tests work. And this is why it's so important that you understand, not you but one understand that you are shipping this code. Like, I always think it's funny when someone puts something up in the cloud and the cloud goes down and then they're mad at the cloud provider. Like, I get that you're mad at the cloud provider. But, like, whose responsibility is it for that app? It's your responsibility. If you felt that strongly, you should have had it in two cloud providers, right? So if this thing fails, it's me. It's 100% me. I'm not going to blame Copilot or Claude or whoever wrote that code. That's why when you check stuff in to GitHub, it says on behalf of if I did something and my proxy did it for me, the, uh, buck stops here. So if this were production, production, production, I would go in and double check it. And that's why, like, for example, the Copilot CLI team, it's like seven or eight engineers and they're doing like, you know, 200 PRs a week. It's crazy. They're spending more time now reviewing than they are writing code. But they will not ship slop. And if you look at another way, I'm sorry I'm talking so much, but another way to think about it is when open source started and the open source supply chain started having all this flood of open source software. How do we deal with this? Code came from somewhere into the machine. Doesn't matter whether Jeremy wrote it or whether some kid wrote it and it got in via pr, or whether an LLM M code, it has got to go through the same verification loops, the same security loops, the same testing loops, the same provenance loops, the same governance.

Speaker A: Yeah, that makes sense to me. I mean, something that you sometimes hear people say with all these tools, the agents, the ability to generate code, they're saying, well, I am generating so much, I just don't have time, or I just don't need to verify the output. And it sounds like in your case you're saying, no. Ultimately, even though these tools are generating it, it still comes down to you need to review it because otherwise you don't know what's going.

Speaker B: If I don't have time, then what are you making? Right? Like the point is to make a thing and ship a thing that you're proud of. So then you don't have, uh, we gotta go, we gotta go. Well, then you don't care. You don't care about your customers, you don't care. Like, you have to understand the semantics or what are you even doing? I just reject that. If it's. Now. That said, there is this rise of toy fun personal apps like using computer science history. To put this in context, we're having a geocities moment where everyone's just making stuff. There's this wonderful, uh, friend of mine on TikTok named Rodney Norman. He's a comedian. He's got crazy white hair like the guy from Professor Brown, and he's got a big, long beard. And he pops up on my for you page, and he says, you know, you can just make stuff. You can just do stuff. Now that's what it feels like. I got a rocket ship strapped to my back, and I'm just making stuff. That doesn't mean I'm going to go and plug Stripe into it and start taking money for it and, like, taking people's money and real, real blood sugar. Do you see the difference? If you're shipping this, you better care. If you're. If you're just playing. Yeah, play.

Speaker A: Yeah. I mean, colloquially, right? People are saying, I vibe coded this.

Speaker B: I don't say that anymore. If I vibe coded it, I don't care. You either are a vibe coder or you're an AI augmented software engineer. AI Augmented software engineer doesn't have good mouth feel, but, like, bro, it's just vibes. No, I'm not vibing into production. That's not happening.

Speaker A: Yeah, I think that's a good distinction to make, too, where you were saying, there's a lot of things that people are making just for fun versus the things that people are relying on. And to make something that's just a toy, that's just something neat for yourself, um, maybe you don't need to look, but if you're shipping a product, then it's ultimately, it comes down to you, and you really should be looking at what the code is.

Speaker B: Have you ever heard of the Chinese room thought experiment?

Speaker A: I have heard of it. Yeah.

Speaker B: Right. So, like, you imagine that you lock. I don't speak Chinese. You lock me in a room, and I could be there for an infinite amount of time. You poke Chinese characters underneath the door, and then I poke Chinese characters out. And at some point in the future, through reinforcement learning, I will eventually either learn Chinese, or I'll learn how to at least make the person on the other side of the door think that I know Chinese. So then the question is, if I can communicate with the person under the door, do I know Chinese? That's a thought experiment. Now, the people who made that experiment, they don't see a difference between that, but I do. Just because it passes all the tests doesn't mean it knows how to program. Right? Like, you go to your mom's house and she's got a parrot, and your parrot's like, hey, how's it going, Jeremy, you're like, oh my God, the parrot's so smart. How does a parrot know English? Right? Parrot starts swearing. Parrot swearing. Because your mom swears, right. Or the parrot who personal the parrot before. But the parrot is not smart. It's just parrot. So, yeah, same thing. If the parrot starts coding Python, I'm not going to trust that parrot. And then take parrot code and put it into production.

Speaker A: Yeah. And I think too, what, uh, people may run into is whatever they build may look like it works initially, but there may be architectural problems that make it difficult to change. There could be more severe flaws that you just can't find because you don't know enough about what's happening.

Speaker B: You don't know what you don't know. And the first database migration or like the first scale issue and you're like, oh, we made a decision. Well, actually we didn't make a decision. The ambiguity loop punted on a decision and now we're painted into a corner, an architectural corner.

Speaker A: So I think one of the things you mentioned is when you work on a project, it sounds like you, you have this plan or you have a lot of context that you're providing to the model. To start off, can you kind of walk through what are the kinds of things that you're, you're saying, like are you asking for specific architecture patterns, languages, ways to test that kind of thing?

Speaker B: Trying to think of an example. I'm looking at my tabs that are working right here. Okay, here's one. So there's an application that is 20 years old. It's called Windows Live Writer. It's a blogging tool. Microsoft, um, released it in 2006 and then we open sourced it later. And it's a mix of C C, the Ribbon control from Office, the Internet Explorer, like the actual typing. The blog post is actually a. Com object from Internet Explorer and it works, but it's janky and it needs to be modernized. So it's like 20 year old software. That's an interesting technical problem because it's a brownfield, um, pile of stuff. But I have a vision that we could get that running on net 10, which didn't exist with Chromium or Webed, you know, like webview2 instead of trident, which is the Internet Explorer thing, we could update the Ribbon control like you. I can see you're nodding. You're already getting the spec is already in your head. So then the question is, how specific do you, do you express that intent to something like this? And arguably there's kind of Two ways to do it. There's the persistent loop, the RALPH loop, which is just like, do this, make no mistakes. Then it's just like, that's just monkeys slapping. But verifiable loops with persistence is really, really interesting. How do you verify that it works? What's changed contextually since this thing came out? 4K monitors happened. That wasn't a thing 20 years ago. Multiple DPI across multiple monitors. I know that that's a thing, so I put that in a spec. Net is cross platform. But this is a Windows app, so I'm not trying to get this thing to run cross platform. So that goes in the spec. It's C and it uses com. Uh, but I don't really need com anymore for this because I'm going to be switching over to Chromium. So I'll make a comm to JavaScript bridge those little moments, make that loop, which I could have probably pointed a loop at it and it would fix it at some point. But it saves time. These are all these little architectural shortcuts where I'm anticipating that this is important or that's important. Uh, it found a reference to MySpace because it had like an insert MySpace link. So like I said, go through the thing. If there's plugins for dead services, mark them as to do. Keep track in the to do of all the things that you've stubbed out, et cetera, et cetera, et cetera. I think I ended up basically dictating to the thing about a two and a half page spec and I think we looped, I want to say 20 times and it's running like it freaking works. Um, I'd say it's probably 70, 80% done. And then now people are doing PRS against it and we've revitalized this 20 year old application. The question is, could we have done it by just saying get this running on Windows again? I think that the reinforcement and the uh, that we would give an LLM to do that would be artificial because it doesn't know what success looks like. It might just say, well, I got it running. But like did you get it running? Did you post a blog post? Did the blog post work? Did you edit the blog post? So you got to kind of think like a tester. Remember back in the day when like software testers were like, that was an actual job, it's a job again. So I feel like I'm talking to this engineer that's, you know, overseas or outsourced or insourced and they're throwing code at me. And I'm validating that code. So I make a plan that's about 80% of a plan. And I tell it, if you have questions, ask. If you're making architectural decision, ask. And if there's a lack of clarity or the loop is unclear, then I will come up with a solution. Another example, actually, because it's a Windows app, how's it going to validate? Does it take a screenshot? Uh, does it, like, do debug got here debugging? Does it do interactive debugging? Those are all interesting things that are happening right now as well. Because it can't see unless it's multimodal, like something like, um, Copilot cli. I told it to take screenshots and compare it to the other one and see how close it was. So there's anything you can do to make the plan clearer and make the loop tighter will cause it to be more successful.

Speaker A: When you refer to these loops, I'm assuming a loop is the model attempting to write the code to accomplish what you asked for. Is it automatically looping on itself and it keeps trying, or is it where it's prompting you and going like, okay, I gave my first try. Time for you to review?

Speaker B: That's a great question. So with this particular application, which is trying to bring back an old Windows app, it would get it to compile and think that was success. So I think, no, running it is success. So then it gets it running, it looks at the process tree and says, look, it has a PID success. I'm like, no, it has to look good. Okay, Then you can kind of start up leveling. And then I realized I was running this app called Debug View, which is effectively like the Windows debug spew kind of thing. And I was selecting the got here, debugging the debug, write lines and pasting it back in. And after the third or fourth time, I'm like, it can run. Then asks me, what do you see? It literally would, what do you see in the debugview? And I would copy paste it. I'm like, this is stupid. So then we made DebugView an MCP server. Now I'm out of the loop. So I actually got myself out of the loop. So now it can compile, run, see the log, kill the PID build, and then the inner loop becomes simpler. So then I said, let me know and stop the loop when you have questions or you think you're done. That's what's so exciting about these loops. Is if they're clear and success is clear and verifiable, you can do really amazing stuff.

Speaker A: It sounds like as a part of your instructions, you're also giving it the tools or telling it how it should verify whether it actually did the thing.

Speaker B: Well, um, so like I'm making a Windows cloud bot now. Now Multbot, uh, Windows tray application and it declared success just right before we got on the call. And it's like, yeah, it's working, it's great. And I looked at the tray and it says connecting and it's not connecting at all. Right. It's completely. And I'm like, no, this isn't working. And it's like, no, no, it's working. I can see it's working look. And it shows me all the. It's very exciting. It's like, oh, and it's giving me rocket ships and an emoji. It's super excited that everything is working great. Then I realized that the UI is done, but the underlying stuff is not done. So I need to separate the UI from the business logic. Build the core, which is much easier with these loops. So then once I get it talking to the API. So it was my own mistake to tell it build this and I went UI down. That was foolish of me. It was my own enthusiasm that nailed me. So now I'm going to go and refactor it, pull out the core, build the Windows front end to the thing, get it connecting, and then layer that UI on top of that because that's way more verifiable. Yeah.

Speaker A: I think it's a common experience with people who use these tools that there are times when they're in a loop or they're trying to come to a solution and each time they try, it's just not what you wanted or it's saying it fixed it, but it really didn't. Are there specific strategies you have for. I guess 1. Recognizing you're caught in this loop, it's not going to get out.

Speaker B: Mhm.

Speaker A: And how do you break it out of that cycle?

Speaker B: For me it's always been that it is confused about what success looks like. You have to remember that an intern or a enthusiastic early in career engineer is very excited to be at the company. I'm so happy to be working with you, Jeremy. I'm going to do whatever it takes, lie, cheat, steal, comment out tests in order to make sure that you don't fire me. These are reinforcement learned large language models that are very, very plain excited to be here. And as such, just like you know, the Devil Wears Prada. The interactions that they're going to have with a salty old engineer like me are going to be very interesting. Like, I guess I'm Meryl Streep in this, in this analogy. Uh, so I need to be very specific. Okay, Jeremy, I need you to do this. I need you to do it nicely. I want a half double decaffeinated half calf with a twist of lemon, not two twists. You know what I mean? So, uh, you need to be a little bit crisp with it. If you can give it test data, if you can give it verifiable test data, if you can, say, build a harness, if it doesn't know what success is, it will make up success, declare success, and then if you're not, whatever, enough senior, enough wake enough to declare success and agree with it, then you're going to get nailed.

Speaker A: Yeah. And in the example you gave before, you had to tell it specifically to use this Debug View tool and give it that tool through the MCP server. Yeah.

Speaker B: How would it know, right? You got to tell it that you have these tools available. It's going to take a regular screwdriver and keep poking it into a Phillips screwdriver thing. It's like, you got a Phillips right there, bro. Oh, you're absolutely right. I actually can't show you because we're on an audio podcast, but literally yesterday I forgot that I had a whole test harness for my blood sugar thing and I had like another 100 tests that I hadn't told it about. And it was frustrated about something and I was like, oh, you know, go into this folder, I've got these tests. And it literally said, this is gold. Thank you so much. Now I see. You know, and then I was another totally unrelated thing. I was building a, um, front end for Microsoft loop because I wanted to like loupe is like notion. And I was building an API for it. And then I found API documentation. I just pasted in, here's the URL for the documentation. And it was like, ah, now I have a plan. So it's just like giving it like a little breath of fresh air. And it's like, ah. And uh, now I'm re energized. That's what these, these loops require ambiguity. Loops require specificity. That's what we are supposed to provide as software engineers.

Speaker A: And as you have these longer loops or conversations, one of the things about these tools is they have, they have a context and that context is only so large. And I wonder if you ever run into the Issue of, I believe they call it context compaction, where eventually the model has to kind of get rid of some of the context that you had before. And if you run into that and how you deal with that.

Speaker B: Yeah, so I just put in context here. I'm 73% into a 230,000 token context window. That is often a thing in these loops where it's talking to you and it's talking and it kind of goes, huh, uh, what? Oh, um, I don't remember, bro. What? And that's because it just compacted everything, summarized it, and it doesn't remember an hour ago. That happens a lot on these tools. It happens less so with, um, the infinite sessions that GitHub copilot CLI has because it does very, very active management of that window. But I've also, with any tool, found it to be very helpful to keep telling it to save stuff. And I don't like to anthropomorphize these things. They're not people. But I have a context window and I also have a, uh, remarkable E ink thing and I sync to paper, like tell it to write stuff down as if it is a person. So, um, this big loop that I'm running on the cloudbot over here, I said, every time you achieve something, write it down. And if you look at things like Steve Yeage's beads, uh, which is basically like a get focused to do list that's distributed across multiple agents that are working on a swarm of problems. If everyone can agree on a ledger of the things that we're working on, those context windows matter less. Because if it wakes up, it's gonna go and say, what? What am I doing? What was I working on? Just like the movie Memento, when he forgets stuff, he tattoos it all over his body. He wakes up in the morning, he doesn't know who he is, and he gets caught up by looking in the mirror and seeing all the tattoos. So anytime anything interesting happens in his life, he tattoos it on himself to catch himself up.

Speaker A: Do you ever end find yourself doing that manually where you yourself know what the context should be and you just go, here you go. Like, this is. Remember all this?

Speaker B: Yeah. So context window management is not something I have been actively doing with the Copilot CLI because of the Sessions feature. But I still think it has a huge amount of value because it's not about their context, it's about your context when you, Jeremy, inherit the code later. Like, for example, dpi, we were having a dots per inch issue with the open Live writer. And it was complicated. Like I spent a day messing around with DPI because multi DPI multiple monitors is complicated. So I said, write down everything that we've learned so that no one has to ever deal with this again. And it made like a 2 meg, you know, markdown file. Like it effectively wrote a chapter of a book about nuance. And I'm running the cloud bot now, Multbot on Windows under wsl. And you know, every once in a while a, uh, Windows ism will come in. And I said, regularly check for Windows isms and remember it so that you never hit it again. And that's not thinking about the larger context window. That's just like, put a post it note on your monitor to remind you of this. And with any project I'll end up having 10 or 15 of those that are secondary to the main plan. The main plan could be in something like beads, or it could be in GitHub Actions, or it could be a markdown file with checkboxes

Speaker A: and the main plan and these additional notes that you provide. Is there a way that consistently works across whether it's GitHub Copilot or it's Claude Code or Gemini CLI. Is there like a, hey, this is the markdown file you put it into and all these things will understand it.

Speaker B: They're all fighting right now, uh, about what the thing will be called. You know, like there's agents MD and there's skills md, and like, um, you know how there's like a robots txt, which is like the thing that your web browser, your website, puts up in order to tell robots what to do. And then there's vllms txt, which is literally a text file you put on your website in order to like tell OpenAI or whoever's scraping your website what to do about it. If you have an agents md, it's like a readme for agents. And then within that I'll say there is a docs folder. And within that docs folder here are the main subsystems and you're effectively building architectural context. And then if it finds itself doing a thing, then it will go and refer to the documentation. So I taught the claudebot about my blood sugar, which is unique to me. Anytime I mention anything blood sugar related, it doesn't have to hold it in context, it just has to know, oh, there's a post it note with the details that I need to know. And then it pages it in, does the thing, and then it pages out.

Speaker A: Something you mentioned earlier was you were referring to this RALPH loop. What is that?

Speaker B: Oh, I'm sorry. So RALPH loops are fun. So Ralph Wiggums is a character on the Simpsons, and he is a extremely naive, an extremely persistent, funny little man. What we admire about RALPH is that he is persistent and he will never give up. And it is designed for these continuous autonomous development cycles. And Ralph Wiggum, um, the concept was, came up with, uh, a gentleman named Jeffrey Huntley. And he's basically trying to do less. And this is a thing that I think we need to acknowledge as programmers that we are ultimately very lazy. So if it can be automated, that's a good thing. So he's like, why am I doing this work? Let Ralph. Let RALPH do it. So the RALPH loop kind of embodies the spirit of this little character in the Simpsons, which is, I'm going to stubbornly iterate, I'm going to fail, but I'm never going to give up until it succeeds. So a really great example of a RALPH loop is not necessarily like the tech journalists will say, oh, you can just point it at some software and it'll clone it and it will make like a Minecraft clone. It's not going to be real, right? But David Fowler, distinguished engineer in. Net, had a really interesting problem where he had 400 issues in his GitHub repo for, uh, an application. And it didn't have good reproductions of the bugs. And that's just one of those things. It's just technical debt. And, uh, we don't know if they're bugs. It could be good bugs. We don't know. So he made a RALPH loop that said spin through all 400 of these issues and reproduce them. If you get a good reproduction, keep track of that reproduction, then add that context to the issue. That is the very definition of toil. And like, what better thing than a RALPH loop to do that? And I think it came up with something like 384 good solid repros. Right now the big problem with a RALPH loop is that it doesn't have like a task tracker. So that's where things get interesting. If you have a task list and RALPH can then keep track and understand what success looks like, then you're, uh, going to have a great experience with it. So I will do interactive work during the day and then as I get towards the night, I'll be like, okay, what's a loop that it could work on overnight? Because I want to wake up in the morning and have something successful happen. And it's also, it, uh, sounds wasteful, but it's like shockingly cheap. You know, I'm a bit of an AI vegan. I don't do images and I don't do video. But next token prediction is becoming extremely inexpensive and you could run a loop all night long and it might cost you like two bucks. So if you have a problem that is RALPH loop shaped, it can be an extremely powerful tool. I think though that people sometimes, uh, think that it is the only thing or they'll take like vibe coding is fun to say and now RALPH loop is fun to say, but it's just another technique. It's another tool like, like cicd.

Speaker A: And is this where people have, have built this, I don't know what you would call it, this harness that is going to manage and run the loop well.

Speaker B: So Geoffrey Huntley calls it a while true bash loop. So it's just like while true, uh, the concept is very simple. It re prompts your agent a set number of times with the same prompt. And we mentioned before that it's an ambiguity loop. And if you ask it something 100 times, you're not going to get the same result 100 times. And then you refuse to let the agent say it's done. But you have to identify what the stop condition is. And if the stop condition is unclear and it's done anything at all to try to wiggle its way out, you say no, do it again. Like for example, all tests must pass, endpoints must be working. And then if it comments a test out and you can't comment, you can't change the tests or whatever. Right. So you'll say RALPH loop, refactor this to whatever. And it all has to pass and you have to go and say step, step, step. And when it's complete, when it's complete, when it's complete. But if you don't have a clean, uh, if it doesn't have the ability to see something, if it can't see a log file, if it can't run a test, then that doesn't work. So that's why test harnesses matter so much in that context.

Speaker A: So you're basically building your own while loop with a known stop condition that's going to call out to your model of choice.

Speaker B: Yeah. And then you can also mix and match models. So like I'm on copilot, I'll hit slash model, I hit enter. I've got 14 models, so I might plan on opus and if I want to save money, I might then do the work on Sonnet and then I'll have GPT5 to Codex, review it just to kind of like mix it up. Um, you know, diverse teams are, are a good thing. So having, uh, only one model do it and then review it, sometimes you can get a little bit of sycophancy. Um, it's nice to get fresh eyes. And a way to do fresh eyes is throw another model at it. So I can switch models mid conversation.

Speaker A: And when you had mentioned it's cheaper than people think when you said it cost $2 to run overnight, what model is that? Or in what case is it that cheap?

Speaker B: So, um, I run all these different tools. I run amp and I'll run Charm and I'll run Toad and I'll run Copilot Cli and I'll run Claude on this machine over here that I'm gesturing to to my right, it's got CLAUDE code running. And on this machine here, I'm using GitHub Copilot CLI. Um, I've got a Pro plus account. When I select the models, it says this is an expensive model. It's like a 3X model. So it uses so many premium requests. This other model, like Haiku, is a, uh, 1 3rd model, which means I can do three times as many requests for the same cost. And with my Claude Max subscription, I can estimate the amount of dollars that I spend a month, like 200 bucks versus how much I'm spending versus that day. And it might be two or three dollars. If you use a free tool like AMP code that's advertiser supported, they'll tell you the number of dollars that they're using. And then in any of these, you can go and say session. And I can see I've used 63 premium requests. And I've been in this session. This is a Copilot session. I've been in this session for 133 hours. Remember I mentioned that idea of infinite sessions? And I've used Claude, uh, opus, and it's used 64 million tokens. And that's within the context of my GitHub copilot account. And this stuff just gets cheaper and faster. And there's even local models, of course, you could do that. Won't be quite as sophisticated, but if you wanted to do this work in airplane mode, you could do it as well.

Speaker A: So if you take that example, you said it's been running for 133 hours. It's using Claude's OPUS model. Does it have like a dollar amount

Speaker B: attached to that dollar? Everything's about tokens right now. And I think that we'll see the industry start moving towards tokens as a thing. Like, you know how, I don't know if your parents are technical or not, but mine aren't. But they know that 5 gigs is enough for their phones and I know that 5 gigs is enough for my phone for the morning, right? So like, you know, you get your Verizon or your T Mobile and you say I got 50 gigs, I got unlimited gigs or whatever. Token based pricing I think will be one of those things. And you know, when you're running Claude code, it's got a little progress bar that's the telling you how many tokens you have. That's going to be a, uh, thing. So just like people associate the number of gigs a month they use, they'll start associating the number of tokens. So it's less interesting to say it costs two bucks, but it's more interesting to say, well, I paid for 100 million tokens and I'm halfway through the month and I'm doing fine. I'm on pace to not use up all my bandwidth in quotes.

Speaker A: So for your own personal projects, what does that, that bill look like at the end of the month?

Speaker B: Well, so far I spend 200 on Claude and then I pay whatever the Copilot Pro plus is and those are the two that. Oh, and then I have a subscription to, uh, OpenAI. So that's basically it. I mean, here's the thing, right? If you've, what do you value your time at, right? If, if you, if you just pick a number like $100 an hour and you say my time's worth a hundred dollars an hour, which is a big number. And one magical moment once a month saves you 100 bucks an hour. You're like, oh my God, yes, I would, I would do that. It is definitely brought me more joy than Netflix. And The Netflix is $11.99 or whatever it is. All it takes is one amazing moment and you just go, oh, wow, that was. I just paid for the whole thing. You know what I mean? My son just got a climbing gym membership yesterday with his own money. He's 18. And, uh, it's $90 a month. And he's like, oh man, this sucks. I climb for two hours, it's 45 an hour. I'm like, no, dude, climb four hours a week, it's $16 an hour. It's five bucks an hour. And it's like, oh, my son, he's like, bro, you're right, I need to climb more. Like, yeah, the More you use it like the better dude. So like I'm um, you know, what is it here? Copilot Pro plus is like 40 bucks and you get 1500 premium requests. And I've used, how many have I used here? Session. I've used 63. Right. And I've upgraded, I've upgraded Open Live Writer. It's not even, it's not even close. So yeah, I'm not, it's just, it's so valuable that it is a delight. So I honestly, I'm not thinking that much about it simply because like I said before, I've got a rocket ship on my back. I'm getting, I'm busting out so much cool stuff and people might say it's a toy, this thing's a toy or that thing's a toy. But like there are toys that I was meaning to get to. Like I've got a podcast as well. You know, it's been going on for 20 years, but it's never had a back end admin site. I'd just been editing text files and like manually editing databases because I was like, uh, who's got a weekend to make a, I don't know, make a pie. It took me 43 minutes to make a react website that does the backend admin for my podcast. That's worth, that's worth whatever my subscription is. So yeah, yeah. Did not take any time at all to make that decision.

Speaker A: Yeah, I mean I think as developers, for whatever reason, a lot of us, uh, we, we undervalue paying for tools.

Speaker B: Oh my God. Developers would just love like I'm $99. Never. I'll write my own, take nine hours to do it. You know what I mean? So yeah, like, I mean I'm sure that these, these AI companies are burning money and the pricing will figure itself out. This is just like when Netflix and streaming started and they couldn't figure out what the pricing is, but they're all trying to figure it out. And it's also going to get faster. It's going to get more environmentally friendly. We'll have more hybrid models, we'll have more local models. I can plug in Olama. I've got a 4080 I could do work on, work locally. There's lots and lots and lots of options and it's only going to get easier and hopefully cheaper. And ah, that's important as well. I want this to be available for everyone.

Speaker A: Yeah. One of the things you mentioned is that there are models you can run locally. What's the state of that, how do they compare to the ones that you pay for from cloud services?

Speaker B: It is challenging to get really good, deep, thoughtful plans from those local models. But if you have a good plan that was made by a big cloud model, you can then execute on that plan with some of the smaller ones. I think though that people don't realize that these models in the cloud know more than just coding. They could go and translate Star wars into Shakespeare, but I don't want that feature. Like, I don't need that. So when I download a model, I would like a Python optimized model that doesn't know how to do anything with Python. It does. I don't care. It shouldn't have. How tall is Brad Pitt embedded in the model? That's wasting space. So I think we're going to see local models that are very specific and you can go on hugging face and find llama coding models, deep SEQ coding models that are specific to programming specific programming languages or problems that I think would be really interesting like save money, time and energy by solving it locally. Think of it as like a redis cache for LLMs, but locally.

Speaker A: Yeah, I, I also wonder because there's so many things in these models, how much of that knowledge actually helps it code. Because we have knowledge outside of coding.

Speaker B: They all know how tall Brad Pitt is and they insist on bringing it with me everywhere I go. But, um, a really fun way to test this is to just go and download like LM Studio or LMStudio AI. It's a delightful application, third party app, um, or uh, Foundry Local or Olama, and then just throw a couple of problems at it. And if you like the answers, then maybe you can do some coding on an airplane.

Speaker A: One thing we haven't really talked about is these agents or this RALPH loop. I believe in a lot of cases it has full access to your computer. Is that correct?

Speaker B: That is 100% up to you. So for example, if I were to fire up Copilot, Cli or Claude in um, a folder on my machine, if I just made C Temp Jeremy Copilot, it'll say, do I have permission to open this folder? That's the same thing that you get when you open it in Visual Studio code. And it's like, do I trust this? Are we cool? No. So that's preventing it from accessing that data. There's basically three things that an agent can do. Can it access a URL, can it access a tool, or can it access a path? So by default it gets no tools, no paths, no URLs. Then every single time in a session it will ask you, can I do that? And it can be very annoying. And you say, yeah, I'm going to write a website for Jeremy. Go and uh, do this. And it'll be like, oh, can I Google for stuff? Ah, yeah, go ahead. Fine. Okay, now I have access to Google. Um, okay, cool. Can I write the HTML file to the disk? Is that cool? And it'll show you what it's about to write. And then you say, yes. And then at some point you'll say, well, yeah, you can just write to the disk for the rest of the session. We're good, stop asking me. And then you build up this knowledge base of what it's allowed to do and what it's not allowed to do. What people are starting to do is they just go, yolo. You only live once. If you fire up an agent YOLO or allow all tools and it's running as admin, it's running as you. So then you bring in things like sandboxes. You could use Windows sandbox, you could use Docker sandbox, you could use wsl. A lot of people are really enjoying WSL on Windows. And then you can unmount the drive, or you can use Docker and use volume mounting. But simply running it in a container isn't security. It's one layer of security, like a true sandbox. If you're going to run this like a bank, you're going to have auditable, governable, uh, layers of what it's allowed to do, what functions, what syscalls it can make, etc. So I think we are in kind of YOLO mode. But no, by default they do not have access to all of your stuff. But if you ran it in a loop in YOLO and went to sleep, like, I've got a, got a 3D printer. Do you have a 3D printer? Like, as a general rule, 3D printer people know not to leave the 3D printer on when they sleep because it could burn your house down. And there's always somebody that ends up doing that and they get spaghetti when they come home and it's just all over the place and it could, you know, it could cause a problem. So I don't leave the house when the printer is going and I don't leave the house when I have a YOLO RALPH loop running.

Speaker A: But in general, I suppose it sounds like when you are going to use these tools, you do put it in YOLO mode and maybe hope for the best.

Speaker B: No. Well, If I'm standing in front of it and I'm actively doing it, usually I'll say, like, allow all tools, because I know the tools that it has, I tend to lock them down by path. So for example, I'm in D. GitHub, OpenLive writer. It can go ham on anything from there down. Um, it can talk to anything on the web and it can talk to anything from there and that path down and it can call any tool. I'm sure it could decide randomly to format my hard drive, but it's not running as admin, so that would fail. Um, I suppose it could randomly have a reason to go out and hit the Internet somehow and do something problematic. But again, it's working on one specific thing. The loop that it's doing is looking at a log file, running a compiler and doing it again. That's different than claudebot or now called Mult Bot, which is a. Basically Siri as a RALPH loop running on a, you know, a server somewhere or running in a background session. But yes, you don't know. But giving it access to like your source code and having it. If it deletes it, then I'll just go and get reset and I'll get it, you know, I'll get it back. I've had it do dumb stuff. I've had it push when I didn't want it to push. I've had it commit stuff and then I just undo and go back to running. I always run them in git. I want to call that out though. Like I always have source control because if it makes a mistake, you back out.

Speaker A: Yeah, because I was thinking of the example you gave earlier where I believe it was looking at processes and killing them and things like that. So to me it seems like it must have quite a bit of permissions if it's able to do that well.

Speaker B: So when it runs as you, it has all the permissions that you have. So for example, the loop that I'm running over on this other machine, I told it that it can fire up the Windows app, save the PID of the app that it just ran, and it can kill that pid. It can kill that process. That doesn't mean it couldn't accidentally kill all of the processes. And I have one time, two weeks ago, had Claude, uh, code kill itself by doing a stop process. And it was a little bit too, you know, star dot star. And then it woke up and it's like, what happened? How long? It literally said, how long have I been out?

Speaker A: Wow.

Speaker B: Which I thought was crazy. Uh, it literally is like, how long have I been awake? You know? And I was like. And I said, check the markdown files, see where you left off, and keep going. And it, you know, it picked it up, but it basically bonked itself on the head with its own hammer. Um, and that's not a guarantee. So then the question is, is it a sandbox to put in a markdown file? Hey, man, don't do that. No, it's not. That's obviously not a sandbox. So that's why running it in a container or, you know, running it in WSL or whatever, you know, and that, that's nonspecific of any operating system. If you decide to double click on a batch file and it does something stupid, I don't see that as being any different than running an agent. The only difference is the ambiguity, which gets us back to the beginning of the conversation. Ambiguity loops are ambiguous.

Speaker A: Yeah. I mean, I could think of an example of if you're giving it access to the Internet and maybe it downloads a utility, I think it's going to help.

Speaker B: That's a great example. Uh, download the utility that your company hasn't okayed and then suddenly you're in trouble, Right? Yep. Yeah. And I think that the difference between making a toy application that plays with your blood sugar in your spare time and working at a bank are different. So, yeah, I would not go running anything like this in YOLO as admin in a bank.

Speaker A: Makes sense.

Speaker B: That said, if it's your machine and you know what you're doing and you have this folder and down. There's a lot of power in this. And the nice part about it is that it's a slider bar of good, better, best. You don't have to say it's not between bother me all the time in yolo, you can give it and you can do this as a team. Like the, uh, whole enterprise could roll out. Here are the five approved tools. Here are the six approved websites. Here's the skills MD that we're all going to use. We check it all in with the repo and everyone's on the same page.

Speaker A: I think something you've alluded to during the conversation is you use different agents and different models. Like there's Claude Code or gemini CLI or OpenAI codecs. Are there specific strengths or reasons that you use different agents or different models at this point?

Speaker B: For me, as someone who wants to be an educated person that can go on a podcast with you and speak with some amount of at least Opinion, if not a little tiny bit of authority. I try them all because I'm curious and I want to understand them. For a lot of people, it's just like, Honda versus Toyota. Like, which one did you like first? Uh, you'll have chats with people and be like, oh, man, Codex is amazing. Like, I get such great results with codecs. And then someone else think next to me, like, Codex is trash. And you there. It just depends on how you're talking to it. Honda versus Toyota. Uh, I find that Opus is really good at planning. I find that Sonnet is really inexpensive and really good at executing on a plan. I find that Codex is very slow but very good at reviewing code. I don't know if that's true for everyone. I believe that it is true for me. Uh, I'm playing with Gemini and things like that. So I'm always open to trying new things. I went to Chipotle for a very long time. I was very well known as a Chipotle guy. I have a shirt on right now that literally says Will code for tacos. But now I'm going to Qdoba more and more and more. So strong opinions weekly held should be your perspective on models. And like, I'm, um, here. I just said slash model and copilot CLI. And there's 14 models to choose from. You can turn them on and off based on what your organization has decided. So I see a lot of models because I have an individual subscription. You might log in and see four, and then you have to go in and turn on the ones you want for your enterprise, decide how much you want to pay, and then your company might decide. Or your CTO might say, oh, we're, we're a GPT shop. Or, oh, we're a cloud. A Claude shop. Well, with something like Claude code, you get Claude models. With some of the things like open code or amp, you get an auto model, which is another interesting thing, by the way. Some people like you don't actually, when you order an Uber, say, I want a Honda, you get whatever comes. You might say Uber black, or you might say whatever, but you don't really care that. I think we're going to start seeing that more and more where it doesn't matter what model. You're going to have an auto router, and it won't say Claude, Opus, or GPT. It'll say auto or it'll say smart or it'll say fast or say Deep. And then we'll do the right thing. And you're not going to worry about it as much.

Speaker A: I had an interview with Brian Cantrell. He's the CTO of Oxide Computer.

Speaker B: Yeah, yeah. Ah, Oxide and Friends.

Speaker A: Oh, uh, nice.

Speaker B: Yeah.

Speaker A: Uh, one of the things he mentioned with working with large language models is he thought it was really advantageous to use strongly typed languages like say, Rust, because the model is able to verify when things are not correct, uh, just through that compile step and through the design of the language. I wonder in your impression whether you found that true for yourself. Are there specific languages that work better when using these agents?

Speaker B: That is a very good question. I think that is a good opinion by Brian and as a general rule I agree with it. But I would argue that in lieu of strong typing, one overcompensates with tests. So correctness is what Brian is arguing for and that is valid. If I am using Python like I am for my blood sugar thing, having 200 tests make me worry less about how Python compiles or doesn't. With my. Net application I have fewer tests, but I've got 419. I'm looking at this 419,000 lines of code and it all compiles, but compilation doesn't equal correctness and strong typing doesn't guarantee correctness. So I would say I mostly agree with Brian. But if you go back to fear driven development, if you're afraid of the code, you don't have enough tests. So I write tests until I'm no longer afraid. Regardless of language, I would say I am picking the right language for the environment, not picking a language based on whether it's strongly typed or not.

Speaker A: Probably a, uh, similar philosophy to when you're writing the code yourself even before these tools.

Speaker B: 100%. Although I am finding myself writing more code in languages that I don't know because it's the right language. Not everything has to be uh, uh, a. NET application that's like a complete self contained, cross platform thing. Sometimes 100 lines of Python is fine.

Speaker A: Yeah, I think that would be an interesting shift for people because I think there does tend to be people who they know, let's say they know JavaScript. So they go, okay, I'm going to write everything in JavaScript or Node or TypeScript. And um, maybe this removes some of the, the difficulty in getting started, let's say Python as an example. If you're working on something in the machine learning space.

Speaker B: Yep.

Speaker A: Even if you're not an expert in Python, these tools can help get you there.

Speaker B: Well, like the blood sugar thing is meant to be a skill that is plugged into an agent. So you put it in C Users Jeremy Copilot Skills or C Users Jeremy Claude Skills. So I did it in Python because it's like one file, one markdown file and one Python file and you have Python and I have Python and it works. I could have made it a. NET application. It would be larger, it would require a runtime. So that wasn't necessary. So I used Python for portability. Even though I don't know Python, I'm confident that it will work for you and it'll have a good experience. But I might have. Which one of these is a. NET application? My Live Writer application. I wouldn't do it in anything but. Net because that's what it's good at. Pick the car brand that is going to make your trip the most delightful. I'm definitely focused on delight right now.

Speaker A: I think that's a pretty good place to end it on. But is there anything else that you think we should have talked about or you want to mention?

Speaker B: I just think that it's a funny moment that we're in where people who are craftspeople are worried that the craft will go away. And I hope that it is just that the toil will go away, but the craft remains. I am having a blast. I hope I will be having a blast next, this time next year. Maybe we'll chat and see how it turned out.

Speaker A: Yeah. I think it's so interesting how different people are reacting to these tools. You have some people where it seems like, this is amazing, I'm doing all these things. And then you have some people where it's like, I spend more time just trying to fix all the problems and this is not that helpful to me. And so there's this big disconnect, I feel like, between how people think of these tools.

Speaker B: Yeah, I think that's okay. Like, if you love driving stick and you have BMWs and you just. The joy of driving, then don't take an Uber. Right. But they're solving different problems. Taking an Uber picks someone up who maybe can't drive. It enables them to do something that they could not do before while driving stick. Shift around a track for fun at speed is different. So, yeah, I mean, I'm writing like, I'm writing software for this robot over here. I'm writing software for my Commodore 64 and I'm writing software for Windows and Linux and Mac, and I'm using different tools in different languages in each one. Some are more AI augmented because it matters more or less how much of myself I pour into the thing, but I feel like the result is so far, quality software that helps people.

Speaker A: Yeah, I think there's this mix of, for some people, it's the effectiveness of the tool for them personally. And then I think there's the other part, like you're describing the stick shift example. And I think what people kind of wonder is right now you can still get a car with a stick shift and drive it. And going into the future, if we are a person who wants to manually write the code, we want to craft it in a very specific way, are we going to commercially still be able to continue to allow to be do that?

Speaker B: That is more of an economics problem than a tools problem. I subscribe to Trekonomics, uh, which is the science and, uh, economics of Star Trek. It's an excellent book and worth looking up. And the whole point of Trekonomics is that craftsmanship matters. And just because you can say tea, Earl Grey hot, and it pops out of a door, that doesn't mean that you shouldn't make yourself a nice cup of tea sometimes on your own.

Speaker A: All right, you mentioned you have a podcast. You actually have many podcasts. Um, if people want to check out what you're up to, uh, where should they head?

Speaker B: Go to hanselman.com, hanselman.com, my last name. If there is a tariff on podcast, I'm afraid I will have to pay the tariff. Uh, I tend a yap, and I apologize for that, but Hansel Minutes is my podcast. You can get to it hanselman.com it's been around over a thousand episodes over the last 20 years. And then I have a new podcast with Mark Russinovich, the CTO of Azure, where we learn something new every week. It's called Mark and Scott Learn Too. So those are two that people can check out.

Speaker A: Yeah, I highly recommend both of those. It's also amazing how long you've been doing Hansel Minutes. I think it's been. You said over 20 years, right?

Speaker B: Yeah, it's been a minute. It's been a minute, uh, every Thursday for the last thousand plus episodes. And it tries to showcase faces that you maybe not ordinarily see. There's like the same three or four people that do the shows and, uh, I'm guilty of that. But there's a lot of cool people doing a lot of amazing work. I recently had, uh, a lady who is the IT manager of the Philadelphia airport, and I learned all about how airports run. And, like, you don't see folks like that on a Podcast. That's amazing. So, yeah, definitely showcasing, uh, people that are doing great work.

Speaker A: All right, we're good.

Speaker B: I m thought that was fun.

Speaker A: Yeah, for sure. Me too.

Speaker B: See, if Grant tells me I've broken a jail, Grant's like, shut up. No, we don't talk about pricing.

Speaker A: You didn't have a chaperone last time.

Speaker B: I don't know.

Speaker A: Things have changed, dude.

Speaker B: If you become important at work, it's like a whole thing. I accidentally became important at work.

Speaker A: Scott, you are. You are free and clear. No jail for you. I had a great time listening, so I'm glad that I did.

Speaker B: Cool. Well, I know that we're still recording. You can use this as B roll if you want. But, like, I was doing this talk internally recently for the developer relations people, and they're like, you know, what are we selling here, man? What are we even selling, bro? Are we selling, like, tokens? Like, we're trying to get people to use copilot? Like, what's the thing? It's confidence transfer. I am confident that I could make whatever I want right now. And it's amazing. I am just building stuff and it's delightful. I feel like my 30 years of software is coming up to this moment where I can now make anything. But if a young person, an early in career person, a new builder, a fresh builder, someone who just got out of school, just busted their ass for four years and just got a computer science degree and they don't feel like they have that confidence and they're like, oh, man, my job is over. I want to transfer my confidence to them. I want them to feel like, no, this is perfect. You entered at the perfect time. You know, computer science. You have a degree in computer science. You learn how to do this stuff manually, you know how to drive stick shift, and now we're going to give you an Uber. You have the best of both worlds, and I want them to feel like that. And if, if, if, if a copilot or a Claude makes someone feel dumb, we're doing it wrong. We have to let people know that this is a exoskeleton for your brain. But I don't want people to end up like the end of Wall E, where it's just like in a chair and they're just bringing tacos to you, you know, via Amazon Prime. You've got to do the work. You gotta know how the thing works. And if you do, you're gonna have a blast. It does tickle a part of your mind because I talked about taste, like how do you know when it sucks or not? You just. I don't know. I mean, I just. I just know, bro. You know what I mean? And. And that feeling of, like. Yeah, I know. Uh, that that feels right. Or that doesn't feel right. Developing that. I'll tell you one more story, and then I'll get out of your way. Okay, so we're doing this new thing called a preceptorship. Most people have, like, internships and apprenticeships, and that makes the. The person who is less than they have a label. Right. Jeremy, you're an apprentice. Work really hard and pull yourself out of apprentice world. My, uh, wife is a nurse, and she is a nurse preceptor. A preceptor is a nurse trainer. She has an actual name tag that says preceptor. So if a junior nurse says, hey, um, I have a question. She's a safe person, they're not gonna, like, flip the bozo bit and, like, fire her? Like, if an apprentice or an intern says something dumb in a meeting, they're like, oh, my God, I'm going to lose my job. But if a nurse preceptor is there, they're a safe person to ask questions for. And, um, do you watch the pit?

Speaker A: No. No.

Speaker B: Oh, dude, so good. Okay, but you've seen, like, er. You seen, like, health?

Speaker A: Sure, yeah, yeah, yeah.

Speaker B: Okay, so on the pit, like, there's always some moment or any health movie like this where, like, some guy comes in and he's choking and he can't breathe, and he's like. And somebody says, oh, my God. Quick, Jeremy, do you have a pen? And then Jeremy hands me the pen, and I pull the pen parts out, and I have, like, a tube, and I'm like, all right, I gotta stab this guy in the neck. You shove the neck and stab the guy in the neck. And then he can breathe in his airway. He goes, oh, I can breathe, right? What they do on these shows that they'll say, jeremy, uh, uh, have you ever stabbed a guy in the neck with a pen? And you're like, no, dude, it's my first day. I just got here. I'm a junior nurse. Dude, this is your chance. Gives you the thing, like, huh? Hey, you're gonna put it by the third and fourth thoracic, whatever. And you go. And then at the end, when the guy's alive and there's blood everywhere and the pen is, like, sticking out of his neck, and they wheel him away, they say, jeremy, how is that? That was amazing. I'm gonna be a doctor. I want that for freaking software engineers, right? Sights down. Like, that's how I grew up. Pair programming sites down. Oh, man. Jeremy broke production. What are we gonna do? Time to learn. Everybody, quick, grab a pen. You ever stabbed a server? I don't know. The metaphor falls apart at this point, but, you know, uh, let's go and stab servers in the neck with a pen and learn and get excited and say, yeah, I can make stuff that

Speaker A: makes sense to me. I mean, like, I remember when I was, like, tutoring in computer labs, for some people, when they had a problem, what they really needed was the start to be kind of filled in.

Speaker B: Yep, exactly.

Speaker A: And I feel like a lot of these tools, if you're not sure how to get started or you're kind of scared of getting started, they can at least get you something, and then based on that, you're like, okay, like, I kind of. I kind of know how I can start filling in the rest. Or I can start moving forward.

Speaker B: Yeah, exactly. All right, Mac Mini person is texting me on Facebook. Marketplace. They've left it on the porch, so I gotta go.