#34 Ralph Wiggum, Tool Search in Claude Code, Cowork, Apple Chooses Google!

Samuel: ...and we're live. So, in case you're somewhere on Twitter, Twitter for us is down. Which includes half our articles, annoyingly.

Kabarza: Give it a refresh. So, we are live. I've heard we're live on YouTube, but presumably not live on Twitter. Just getting an X now.

Samuel: I'm going to try the VPN. Let's see where Elon is. Is it live?

Kabarza: Check in. San Francisco, wherever he is. See if it's live there. So, you use a VPN, do you?

Samuel: Maybe that was not the smartest choice.

Kabarza: Why?

Samuel: I used the VPN and I was gone for a second.

Kabarza: Oh, I see. Now you're back. Wait, it seems like my X is working. Can you turn down the volume of your headphones because I can totally hear myself.

Samuel: My headphones? That's my mic probably being very sensitive. I've thought about getting a new mic. Okay. So, turning down the volume. We got three people. I still hear myself. How is everybody doing? You people, just do us a favor. Three people, are you on Twitter? Because I can just see one person on YouTube, so there must be two of you on Twitter. Let us know where you are. We can't check where you are because if you let us know on Twitter, we can't see you. So, I guess if you're on YouTube, drop a comment.

Kabarza: Oh, Twitter's coming back. Twitter's back. Okay. I think the podcast kicked it into gear. I think we were the ones that brought Twitter back.

Samuel: Okay, let me refresh all my pages. Yes, yes, yes. We're in business. Just preparing all my pages.

Kabarza: I'm also preparing here. Putting myself in the center of the camera so you have an easier time editing.

Samuel: Oops. Oh, yes. Yes, yes, please. Thank you. Am I changing color?

Kabarza: Yeah, because of touching the HDMI cable. Right, we're starting to see people pour in now. Hello everybody. Cheers for the like on YouTube. Drop us a love, drop us a chat if you're hyped. Have you used that new hype button on YouTube at all yet?

Samuel: No. I'm too scared to press it. I don't know. I just see it on some videos that there's a hype button. I'm too scared to press it in case it draws attention to my lurking.

Kabarza: Maybe I haven't seen it. I mean, I live in Europe and...

Samuel: Yeah. You don't get a lot in Europe, do you? Because you're scared of everything. I'm in Britain, so we're not much better.

Kabarza: Well, we have, you know, those caps that you can't remove from bottles.

Samuel: Oh, yes. Oh, yes. Yes. Did you want to tell us about your site of the week?

Kabarza: So, I actually have a few things to show. Let's just start with that.

Samuel: Okay. You're going to start. I'm going to go get myself a beer.

Kabarza: Go solo. Fly solo, butterfly. So, to everybody who's watching, have you seen this? You probably have. Are we cooked? Well, some people in a specific industry are probably cooked. People who might need to get a job. Well, that's quite funny because that includes us as well. But I'm not talking about us software engineers and designers. Yeah, this is insane. I don't know if this is real-time or a recording, probably not real-time, but it probably will be real-time if it's not. And then imagine you have a call with somebody... who can you trust? No images, no videos. And now this...

Samuel: We are cooked.

Kabarza: Well, I said not we, because we don't need a new job because AI is not taking our job, right?

Samuel: What are you chatting about? AI taking some people's jobs. Oh god. You still beating that drum?

Kabarza: No, this specific one. Have you seen this video?

Samuel: No. It's so impressive. Oh, I think I've seen something similar with like cartoon characters, but I mean, this is realistic. Like really realistic. Look at the hair.

Kabarza: That's what I asked. I don't know. But if it's not, it will be soon. This is insane.

Samuel: Is this real latency or pre-recorded? So, is this recent then? Because of course Levels.io released this like one of the first ones.

Kabarza: Oh, yeah. There he is in the bottom right. So he's made it into a video thing. There he is. He looks fit as a girl. I'll send him this clip. It's crazy. Yeah, it's crazy how he's kept his shirt and everything.

Samuel: I think the way it works, you can create a character that you want and then you take a video of yourself and say make this, basically make me into it. He said made with mocap, motion capture, on my app.

Kabarza: Shout out to Higgsfield who wanted to sponsor me for $30 for a dedicated video.

Samuel: No way. Yeah, joke. Oh man, that is shameful on their part.

Kabarza: Anyway, so this is an old website, but maybe he's added a video feature to it. Anyway, cool. I thought you were showing a website. I was waiting for you.

Samuel: Oh, okay. This is just warming up our audience with fake e-girls. Okay, great.

Kabarza: So, this video this week went quite viral with almost 30 million views. It's about words per minute. How many words per minute can you read? And I saw this and was like, what if I make this into an app? So, I prompted and this is what I have. If I reload, you see here we have a text box, word per minute, the anchor letter, and I can just start, and then I have a one-second pause at the end of each sentence. I think it turned out pretty cool.

Samuel: Yeah. So, you can set the words per minute and just build it up. So, I can start and let's see if it actually real-time adapts. Yes, it does. I just added a prompt saying can you make this update the animation engine real-time.

Kabarza: And what's the importance of the red letter?

Samuel: It was in the design and I just took it, but I believe it is important to have the words in a specific anchor point. I'm not sure what the formula is. It looks like it's the first 40% of the letter maybe because we don't read the entire words. We just see it kind of like as a shape. Maybe that's the start and the end of it.

Kabarza: We never read the word. That's why it's also really difficult to notice mistakes in a word. We see the start and the end of each word and that's how we read it. I suck at this, my brain trips over itself. Maybe my brain reads it but I can't say it. I really struggle with fast speaking.

Samuel: So with this I could turn it into an app to train. What was your initial prompt?

Kabarza: Uh, my initial prompt. Let's reveal all. Create a GSAP project. It is literally as simple... okay, I was talking while doing something else. Anyway, and then we can do this read-per-minute kind of project. I'm basically saying that we need controls to move forward and backward, which it didn't do at all. I was actually in Cursor, using the voice input.

Samuel: I was going to say you should run your prompt into an AI to help refine it.

Kabarza: I do this. So anyway, we had the prompt and I was quite frustrated because typically I build the HTML/CSS in Webflow for more control. Here I didn't. I just wanted to see if it's possible and then I got really frustrated because I wanted to change some structure in the HTML and CSS and I wasn't comfortable doing it literally in the code. I was then prompting CSS/HTML changes, which I don't like. Anyway, so that worked pretty well, I would say, after about 40 minutes of work. I haven't checked the code at all. We'll get into that, but I think we're beyond the point of checking every line of code. Vibe coding is now how we're all coding.

Samuel: Exactly. This is very interesting. So, what I did, I have an idea for an app and I thought I will talk to ChatGPT to create a markdown file to plan everything ahead. The app is simply to order cocktails at my place. When my friends come over, they can scan a QR code, the app opens, and they can see what drinks I have available and they can order. It needs inventory management, CMS, authentication, commenting...

Kabarza: Basically what a bar would have, but probably better and actually works.

Samuel: Yeah. And I have a good list of specifications. I was thinking if I should do it live, basically start vibe coding it.

Kabarza: That would be really fun. Yeah. And we could show people how to think about it. You are a senior software developer. You can...

Samuel: I'm a senior AI engineer. I've been doing AI engineering for six years. I've got six years under my belt. To be fair, I was playing around with AI back when ChatGPT first came out.

Kabarza: You started very early on adopting it. Let me say, we had this project, and you created a prototype for the client and it was really impressive. I saw the prototype he made within a week and it was an MVP. The UI was clean, the UX was thought out, there was backend, authentication, everything.

Samuel: Yeah, I would be interested in creating this with you. Maybe I bring the MD file, say hey this is the idea, and then we see how long it takes us to create it.

Kabarza: Yeah, I think you'll be surprised. What do you want to use? Cursor, Claude, some CLI?

Samuel: I'm thinking what would help people, like making a tutorial out of it.

Kabarza: Well, if you care tremendously about design, the vibe code thing is not really your avenue. Designers don't really belong in this world. You should be willing to prioritize the product over the interface. If a designer wants to jump into vibe coding, you need to relinquish your ability to drag and drop and learn how to create design with code. Even I'm not going to do design through AI-assisted coding. I'm going to accept what AI gives me or give it rough guidelines.

Samuel: So we do something that can start from zero. If people can watch it and be like, "Okay, I have a vibe, I have a code idea, I want to vibe code."

Kabarza: Let's do it. Let's do something next week. Anyway, let's crack on with the show because that sounds really fun. Speaking of live, let's go through the intro. So this week, Apple accepted defeat in the AI war. We will talk about a paradigm shift based off Ralph Wiggum, which is quite funny. What happens if you let AI code for a whole week uninterrupted? And some Anthropic drama. All of this in today's episode. This is Command AI. I'm Kabarza.

Samuel: And I am Sam and this is a week in AI, design, and dev.

Kabarza: I would say we need to change that to "all of this and more" because there's more we're going through.

Samuel: Nice Tom. I remember you, Tom.

Kabarza: You remember Tom? This is so funny because Tom was at the table when we came up with this idea for this podcast. Tom, do you remember us talking about it in the food place at Webflow Conf?

Samuel: Super cool. Yeah, let's go with Ralph Wiggum. I can't wait to learn more about this because I have some questions.

Kabarza: Okay, cool. Right, I might butcher this, but we're going to go for it. So, we're stepping into a new era of software development where the latest models are no longer relevant. And how we get ahead is the way we orchestrate AI, and it all begins with the unlikely Ralph Wiggum, the famously clueless character from The Simpsons.

Samuel: Here's what that means and how it actually works. It all starts with Jeffrey Huntley here in this very trippy video. You really need to go in and watch the source to get it.

Kabarza: What is Ralph?

Samuel: So, Ralph is a new coding technique. It's less about the models, it's less about the tools. It's an infinite loop that will run tasks infinitely until it gets the job done. It solves the problem of context. When you're in Claude Code, with every message, you're adding to the bloat of that context. What Ralph does, it sits outside Claude Code and gives it the prompt with a fresh context. Everything is brand new. And then it writes the results into an external file to keep track of what's been done. With every single prompt, it's brand new because AIs get dumber the more context you add.

Kabarza: Okay, so here's a for loop here. This bit of code will run infinitely, or at least in the max iterations, and it will take the prompt and run it basically. It will run this prompt into the AI and take the output and put it back into the external file.

Samuel: And the beauty of this is that the idea of running multiple agents in parallel is kind of solved with this because you're giving Ralph autonomy here to decide what the next appropriate task should be. It orchestrates itself, outputs that to an MD file, and then the next agent will take that and run with it. There'll never be crossover.

Kabarza: You mentioned that you're already kind of doing something like this.

Samuel: Yeah, this is interesting. In Cursor, I have multiple chats. One chat is for styles, another for animations, another for the core animation engine. I separate them to manage context because if I mix everything, it gets confused. And I typically have an MD file where I ask the model to summarize its learnings and mistakes, so I can give it to another AI model. That MD file for me is kind of like memory.

Kabarza: The external memory is really crucial to this. You're almost saying "advise the next agent." The important thing here isn't Ralph, it's the fact that we are now entering this new age where we were once fussing about what model to use. That conversation is behind us. We're now learning what it means to be a software developer using AI. You will no longer be hired for the code you write because that's a commodity. The important aspect is how you orchestrate, how you engineer, understanding how all these puzzle pieces fit together.

Samuel: You mentioned a character from Rick and Morty.

Kabarza: Oh, yeah. Mr. Meeseeks. You ask Mr. Meeseeks to do something, it comes to existence, it does it, and the moment the task is done, poof, he's gone. The whole joke is what if you give it a task that it can't do, and that's when it starts to ask for another Mr. Meeseeks to help, and another. That is kind of like the sub-agents we do. But it's interesting because it will try until the job is done, and the next Mr. Meeseeks is also kind of like fresh context.

Samuel: So what do we do with the coding for one week thing?

Kabarza: Yeah. I'll show it. So here we have an article by Cursor where they gave a task to create a browser completely from scratch. They gave it a whole week to run uninterrupted. They wrote this article about it, and it's pretty interesting. You see here how it starts to render HTML and it doesn't really work at the start, but as it iterates, it gets better. They say it kind of works, but it's not perfect and has lots of bugs. It wrote like three million lines of code.

Samuel: What they are doing, they have multiple agents doing different tasks, right?

Kabarza: Yeah, they don't mention Ralph at all. I think this is more of an exercise of how far they can push AI. What they say is interesting: "Multi-agent coordination remains a hard problem. We still need periodic fresh starts to combat drift and tunnel vision."

Samuel: I think they've encountered all of the things that Ralph solves. They are working underneath AI, whereas I feel like the Ralph thing works above AI. It's so outside the box that it rewrites all of these issues.

Kabarza: Time is the concern here, I believe.

Samuel: Time is not an issue now because you can build features in minutes as opposed to days.

Kabarza: But if everybody can do it, the whole industry is changing. There was this idea that we will have billion-dollar companies this year made by one person.

Samuel: You can have a lot of Mr. Meeseeks. A lot of Ralphs riding around.

Kabarza: So, I've just released a video on another groundbreaking moment, which is the tool search tool in Claude. If you don't know, it's a way of handling MCPs, which take up a lot of context. Tool search just takes the description of the MCP and lazy loads it when it needs it. It's now been released on Claude Code. The MCP layer is completely empty. You don't need to do anything, just update Claude. It works in the desktop app, which is really handy.

Samuel: So now you can add this in Cursor. Whilst you're there coding, little side quest, add a thing to our little page that we've got for our news articles. Jobs are good.

Kabarza: I understand the idea, but I didn't have a proper use case until this week. We launched a website for a client and the images are not how I want them. I want to use the Webflow MCP to download all the images, change them, and then re-upload them.

Samuel: So should we move on to Apple Gemini?

Kabarza: Yeah. Have you guys seen that Apple has given up and they are now resorting to Gemini to run their AI-powered Siri? Apple Intelligence is sort of Google Intelligence.

Samuel: Yeah, basically. They say we're using this foundational model from Google for our foundational model.

Kabarza: Here's my take. People are focused on Gemini. Google has an open-source coding AI called Gemma. I think Apple Intelligence will be built on top of Gemma, and they'll be able to call it whatever they want. I don't think it will be a direct copy of Gemini. Apple doesn't manufacture anything, they just design. They take a base construction layer and add their own layer on top. I don't think this is all that controversial.

Samuel: Everything is a wrapper. It's just changing the layer of abstraction. I did see Apple was planning to pay about $1 billion a year to utilize Google AI.

Kabarza: I wonder what's going to happen to OpenAI because right now if Siri can't sort your issue, it'll forward it to ChatGPT.

Samuel: It is what it is. It's the cost of doing business. It's funny how one arm of a company is taking another company to court, yet you've then got handshakes and smiles in other areas. It's just business.

Kabarza: I can show my screen for Claude Co-work. Have you seen the video?

Samuel: Of course.

Kabarza: So there is a new tool by Claude. It's called Claude Co-work. It's going to help you with any tasks you have. You see, create a file, crunch data, make a prototype, organize files, send a message. It works with your Finder and file system. It can access your desktop and do everything. It got almost 50 million views.

Samuel: In summary, I would describe this as Claude Code for normies. I can already do this all with Claude Code. Yesterday, I had all my hard drives out, and I use Screen Studio to record my videos. I wanted to move the Screen Studio file into the folder associated with the video. I told Warp, my AI terminal, to look at the name of this file, do a fuzzy search inside this folder, and if you find something, let me know, and then I'll decide if I want you to move it. It wrote a bash script and went through it. Co-work is not for me.

Kabarza: Interesting. I understand. It further exemplifies the seismic shift we're seeing, not only in software development but also in normies, where we're starting to think about what systems we can build to get micro jobs done. We just want to clean our desktop, how can I do that more efficiently? And did you know they built it in a week and a half using Claude Code?

Samuel: That is insane.

Kabarza: It's really cool. I made a video a few weeks ago and I asked my brother, "If you could build any app, what would you build?" And he gave me an idea. It was beautiful to give someone that power to build their own software to get stupid little things done.

Samuel: Thinking do we cover the Anthropic blocking apps? My take was always going to be I sort of see both sides.

Kabarza: I don't know what happened.

Samuel: So users of Open Code, another popular CLI, got a message saying their credential is not authorized for use with other API requests. Open Code allowed you to use your Claude Code subscription with their tool. Anthropic is attempting to force you to use Claude Code if you're on that max plan. Dax, the guy from Open Code, said quite rightly that models are not sticky, it's the ecosystem around it. On the one hand, Dax understood; on the other hand, a lot of people are very frustrated. All I would say is go watch their videos because I'm sort of in the middle.

Kabarza: If there is a loophole, then people can use that loophole.

Samuel: This one is interesting. OpenAI has acquired Torch, a healthcare startup that unifies lab results, medications, and visit recordings.

Kabarza: On the one hand, you've really got to be careful with your health. The fact that I need to say that means there are plenty of people not realizing that when they give their health to AI, it could one, hallucinate, and two, be composed of answers from losers on Reddit who don't know what they're talking about.

Samuel: Yeah, it's like parents taking advice from Instagram Reels. They're exposed to short-form content and everybody is talking about something they have no idea about. Now AI is learning from that.

Kabarza: I think it's less about hallucinations. It's just about the information that's been given is not necessarily grounded in truth. It's grounded in Reddit, which is a massive source of information for OpenAI. It feels like nobody knows anything and there is less and less care for knowing specifics. Everything is an approximation of truth, but truth cannot be just approximately true.

Samuel: You've hit a real pet peeve of mine. We've become so accustomed to "trust me, bro." I was on Instagram and I get this story about the basis of my favorite band, and the person just ends the thing. Why are we in this world where this person feels no desire to say "Mark the bassist told this story in an interview with Rolling Stone"? Why do they feel no need to cite their sources? My overall message is question everything. Don't just accept things at face value.

Kabarza: To be said, who knows what Torch and OpenAI are going to do. The worry would be that it would take Apple Watch data and your health data and give you interpretations of that, which at the state of AI currently, is worrying.

Samuel: Yeah, take it with a pinch of salt. Always. You know the joke, you go to Google and Google is like, "You got cancer."

Kabarza: Code is cheap, software isn't. Theo released this video. It's something we've spoken about a few times. The interesting thing here is that it's expressing a lot of the ideas that we've already spoken about. I've been talking about this idea that SaaS is dead because we used to look to SaaS products to provide us a service, whereas now you can build a tool in no time at all.

Samuel: Theo and Chris's arguments differ from mine in that they actually think that software is still going to be expensive because although you're not paying for the lines of code, the real engineering lies in the abstraction and the architecture. This is where we need to lean more into. I do think you need to write things from scratch more and more because I cannot see a world where people are leaving university only having touched AI to have the intricate knowledge to understand how systems are put together at an enterprise level.

Kabarza: Last but not least, Apple Pro Apps enter subscription model. This is a new thing that I almost missed. Apple Creative Studio, putting everything under one subscription.

Samuel: Yeah, Final Cut Pro, Logic Pro, Motion, Compressor. You can still buy these things individually. Keynote, Pages, and Numbers are still free. However, there will be things you get when you pay for the subscription service. I think they've balanced it quite well. Students and educators get it for $29 a year, which is crazy.

Kabarza: Comparing it to Adobe, which is going to cost 80-something euros per month for the Creative Cloud.

Samuel: The backlash is simply subscriptions. And we've heard mutterings about this for a while with Apple. But I do think they've done it well. They've given you the software, they're still going to update it, and maybe we'll get features later, but the subscribers will get the latest things first. Totally happy with all that.

Kabarza: Okay. So, Webflow, after a long time of not making any videos, came up with this ad. I personally find the ad to be funny, but corporate funny. What do you think about the video, but also about the whole idea of websites and AI accessing your website?

Samuel: Yeah, the video is great. They've been speaking about what they call AIO. Instead of optimizing for SEO, you're optimizing for AI crawlers. What is their obsession with it? What am I not seeing?

Kabarza: I think it is because we have less traffic coming from search engines. More people are asking Google AI, so they're not coming to your website as much. So you probably want to communicate that with LLMs. Webflow now has an LLM file, similar to a sitemap, that they can take. But I agree, if you make your content good, it doesn't matter. If you use best practices and make it accessible, it's going to be accessible to these crawlers as well. I feel like my usage of Webflow is getting less and less because I'm just using Cursor more.

Samuel: It's difficult because there's nothing new to learn.

Kabarza: We are teaching, and I'm sure you've seen a lot of people are not like you and I, sitting and just researching and learning by reading documentation. A lot of people want to watch a tutorial and follow along. I don't know what happens when there are fewer tutorials.

Samuel: I think there'll always be a medium for watching someone do it, which we'll hopefully try and organize next week. We're going to try and do a live AI coding session where you're going to build a bar ordering app.

Kabarza: What should I prepare?

Samuel: Just know what you want to build in your head. We'll start from scratch. We'll announce it on Twitter and YouTube.

Kabarza: Cool. All right. Well, that's all we have time for. Make sure to like, follow us on Instagram, on YouTube. Join us next week for all the news in AI, design, and dev. Let us know in the comments if anything you've heard or anything you want us to cover. I've been Sam.

Samuel: And I've been Kabarza. Keep on vibing, baby.

Kabarza: Peace out.