# Fable 5 and OpenAI's Curious ChatGPT Update

## AppStories

### Monday, July 13, 2026

**John** [0:11]: Hello and welcome back to another episode of AppStories. Today's episode is brought to you by Daylite. I'm John Voorhees and I have Federico Viticci with me. Hey, Federico.

**Ticci** [0:20]: Hello, John. How are you?

**John** [0:21]: I'm doing really well because you know what? I've got a surprise for you right off the bat, Federico. good surprise. I think you're going to love it. You know, we've never done AppStories merch. I guess we did stickers at one point, but for the most part we haven't. And you know what? We've never done a t-shirt in particular. So I went out and I found the best AppStories themed t-shirt I could find anywhere in the United States of America.

Okay. Let me just reveal the shirt to you.

**Ticci** [0:26]: A

**Ticci** [0:49]: Okay. App State?

**John** [0:53]: App State.

**Ticci** [0:54]: So if you're not watching the video version of AppStories, John is wearing a t-shirt that yellow and state in white. black t

**John** [1:02]: On a-shirt.

**Ticci** [1:03]: What is an App State?

**John** [1:04]: So App State is short for Appalachian State, which is a North Carolina state college here. And it's located in a town called Boone, North Carolina. And one of my, some of my relatives have gone there before, my brother's kids. But it is called App State. This is the flag I sent you for the Mountaineers. Remember that? I was in Boone. I was in Boone over the weekend. I'm going to talk about this a little bit on Unwind.

But I was in Boone and I walked into this college t-shirt shop. You know, all college towns have this where there's like sweatshirts and t-shirts and all the stuff you can get with the school logo on it. And I saw this t-shirt that said App State. And I thought, you know what? That's perfect for AppStories. So that's why I'm here wearing my App State t-shirt.

**Ticci** [1:54]: That is actually pretty perfect for the show.

**John** [1:57]: I thought so too.

**Ticci** [1:58]: Yeah. Thank you, John. Thank you.

**John** [1:58]: All right. So enough about my wardrobe, Federico.

**Ticci** [2:02]: I mean, we can talk about other stuff, maybe, maybe it's not a good idea for, for AppStories. What is the topic today, John?

**John** [2:04]: but

**John** [2:10]: Today's topic is the latest rash of models that have come out. We're going to talk a little bit about Fable 5 because you and I have had, I don't know what it's been like a week, week and a half to use Fable 5 here and there because there are a lot of restrictions on it. And then just as right before we recorded this about a day before, OpenAI released its new series of 5 .6 models, Sol, Terra, and Luna.

And we've had a little bit of a chance to use those and also the app that went along with them, which is now, it's still called ChatGPT, but it's a very different app than the ChatGPT that many people know and love. So we're going to talk about all those things. And I don't know, where do you want to start? Do you want to start with Fable 5 since that's what

**Ticci** [2:41]: Yeah.

**Ticci** [3:01]: we've

**John** [3:01]: used the most?

**Ticci** [3:02]: Yeah. Tell me how you used Fable.

**John** [3:04]: Yes.

**Ticci** [3:05]: I think we both built very different things in very different ways.

**John** [3:11]: Yeah, we did. And for me, I mean, first of all, the way, you know, Anthropic is allowing subscribers to use Fable 5 is that you could use up to 50 % of your token allocation on your subscription for a week to build things with Fable 5. Now they've extended that since then and they've done a reset of token limits. So it's ended up being a little more than was initially anticipated, but there was like a hard cap on how far you could get with your 50 % allotment. And where I started was with a project that had been sitting on the shelf since February.

And I mean, listeners probably remember like, I would say Federico and I got more deeply into using things like Claude Code and then later Codex when it was released in the holiday season, you know, the winter holidays around Thanksgiving, Christmas, New Year's in that period. And so back in February, so not too long after I had been doing this, when we were still using Claude Opus 4 .6, I had the idea of building something that I've always wanted. I mean, all the way back to 2015, when I was first writing at MacStories, which is a track changes viewer for markdown files that are version controlled with Git, which is a very specific MacStories thing.

Because, right, I think of all the apps or all of the technologies that we've used at MacStories over the years, one of the ones that has endured the longest is Git and GitHub. Because what we do is markdown file or text files, just like code files. And so we have a repo where, especially for like the club, for club MacStories, when we do our newsletters, where each of the authors writes their piece, commits it to GitHub, and then other people can sync that locally, look at it, comment, make edits, and then sync that back. And then at the end, when we're all done editing everything, all those documents are on everybody's computers, and they can just be assembled into the newsletter.

And it's a really pretty efficient, asynchronous way to edit things. It's not, you know, real-time collaboration or anything like that. It's each person working individually when they can, syncing it, having it picked up by somebody else, and then finished. And the one thing that always drove me nuts about that was that Git Diffing tools, which are made for developers, are just not good for pros, not good for the kind of articles that we write. You know, there are a lot of good tools for developers out there, but they're too complicated for plain old to text. And they don't visualize the changes in a way that makes sense from an editorial standpoint.

And so, you know, I've tried a lot of them. I've tried to Git Tower, Kaleidoscope, all these things. And they're, you know, some of them are okay, but none of them work great. And so what I wanted was something that would allow me to see those changes and see them in a way that made sense from a writing standpoint. I wanted to see things that were inserted, things that were deleted. I wanted to see punctuation changes. And I wanted to be able to filter by all those things because, you know, for instance, if somebody adds a bunch of commas to your document, it's very hard to see that when you're looking at a bunch of track changes.

But if you could filter just to the commas and have an easy way to navigate from one to the other, that would be great because then you could just see what was done without a lot of effort. And so that is something back in February, I started prototyping with Claude to come up with a solution that would work for writers for punctuation and grammar and all this stuff, all the changes. And it worked. And I could tell that it was doable.

But it's a very complicated UI to really get right so that it looks good and is readable at the same time. Because you don't want to like peel away too many of the changes, make them too vague so that they're hard to track what exactly has changed. But you also want it to remain a readable thing. So I worked on it for a while.

Claude got really stuck on the UI, just couldn't come up with something that was good. And so I shelved it. And so that was the first thing I dumped into Fable. And it was a because this is something that I fiddled around with for quite a while, prototyping it in February. And when I dumped this project into Fable 5, the only thing that slowed me down was hitting my five hour token limits because I couldn't do it in one five hour shot allotment. I had to do it over like two or three.

And so Friday night, I started, I got as far as I could before I went to bed. And then the next morning I got up early and finished it up. And that took, you know, so that was over the span of maybe 12 hours, most of which I was asleep. And then I handed it off to you because you had for already for MacStories, you have a landing page for all of our internal tools that you built with various

**Ticci** [8:34]: agents. It's a whole system that I set up months ago. And it's a private space for us based on the sort of the same reusable foundation for multiple MacStories users. And it's all a collection of internal tools that we use every day, including uploading images to our CDN files to our CDN, summarizing Apple press releases, searching my iOS reviews, which I'll talk about in a minute. But yes, it's a private space for MacStories with actual tools, like all the images that we've been uploading for the past two months, almost. They have all gone through this tool that I built based on this common foundation for all of these internal tools.

**John** [9:21]: Right. And one of the tricky bits here when I was adapting my thing for our internal tools is that I built it as a prototype that was meant to run locally on my Mac mini server. And we have these tools on a cloud server. And so what Federico did is he sent me, he sent me a couple of the projects, you know, I cloned the repos, had Fable 5 look at how it was implemented along with the instructions, you sent to me, and had it implement that.

Now, it didn't do a perfect job. I didn't have, I think, all the context that Fable 5 needed from my end. But it was close enough that when I told Federico on Saturday, look, it's ready, see what you can do, there were some bugs that had to be worked out. But by the afternoon, my time, it was finished. And some of it you even did on your phone using cursor, I believe. I mean, this was like, over the course of just over just under 24 hours, where either most of that time I was asleep, or you weren't home, or I wasn't home. And we still got this all put together pretty quickly.

**Ticci** [10:32]: Yeah, yeah. So from my end, I, so I knew that I want, so I, let me start by saying that I left a couple of projects intentionally open for Fable. I've been using Codex as my main agent and sort of system for building stuff for the past few months. And spoiler, I am, I have returned to Codex after my experiments with Fable. I, this is a whole separate conversation that we can have later in the show. It's just more, I just feel more comfortable and I just have a preference for OpenAI models. That said, I think it's undeniable that Fable is a model in a completely separate class in terms of model size. And this is something that we can talk about when we compare Fable to Sol, for example.

**John** [11:25]: Right.

**Ticci** [11:26]: And it has more quote unquote, taste for front end design, application design, mobile UI. Like it, it has a leg up compared to GPT-5 .6 when it comes to design in general. So I intentionally left a couple of tasks in my project waiting for Fable to become available. One of them was adapting your editor tool for different markdown drafts in our GitHub repos. And the other was a, this project that I've been teasing on Mastodon called MacRemote, which is a way to turn your, your Mac mini or Mac to the server and the local CLIs or the local MCPs that you have into, into actual like publicly accessible MCP servers with authentication and permission system.

And, uh, the UI of the MacRemote host app on the Mac was kind of terrible because GPT-5 .5.

**John** [12:23]: a

**John** [12:32]: Right. Because as

**Ticci** [12:33]: designed it Codex and it was horrible. And I wanted to spice it up a little. So the first thing I did with the Claude app, uh, on, on the Mac, uh, using Claude Code in the Claude for Mac app, was redesigned MacRemote completely. And while Fable was working on that, sort of, uh, also do an in-depth review of the whole code base to find any lingering issues or bugs. And so that ran overnight. So I, I let it run overnight with UltraCode and, uh, I woke up the next morning after six or seven hours and it had completely redesigned MacRemote launched it, inspected it, rebuilt the app a bunch of times, closed a bunch of bugs.

And, uh, obviously, thankfully there were no security related bugs. So once again, confirming that, uh, GPT models are really good at that sort of stuff, but there were a couple of like, uh, issues with integrations with local CLIs and stuff like that.

And so I woke up to a really, really improved, uh, Fable designed an entire onboarding flow, uh, did research with the Apple human interface guidelines, uh, what's appropriate, what's a, what's like, uh, a subject, what are suggested patterns from Apple when it comes to designing an alert window, did research on a bunch of blogs, including MacStories for understanding what's a good onboarding flow, what's a good tutorial to teach users about using a Mac app. So that was really cool.

And I went back through the reasoning, um, steps and I saw that I did research on, on MacRumors, a nine to five Mac, uh, downloaded and analyze screenshots of popular Mac apps, the research on Apple design award winning apps to understand why are these apps considered good? And how, like, it was really, really remarkable the extent to which it went to understand the goal. And my next task was, uh, I wanted to, uh, I wanted to add a fancy animation.

So when Mac remote connects, uh, it shows you a connection screen that says you are connected. You have opened a tunnel on your Mac that exposes these tools with these permissions to MCP clients. And I wanted to have a fancy animation for that. And so I had this idea in the back of my mind, like when Fable comes out, I'm going to have design a metal based, uh, GPU accelerated shader that animates when, when a connection is open and the tunnel is open on your Mac. And it did that.

So I had, I had to design this metal based, uh, shader that creates designing this fancy sort of glassy morphing animation

**John** [12:47]: um,

**John** [15:24]: that,

**Ticci** [15:25]: Sort of just moves around, bounces around the screen when Mac remote is open. So that was really cool. And fun. Finally, uh, and again, this is before I moved to cursor. I also took your project, uh, which you sent to me and I cloned the repo that you sent me. And I said, Claude, this is a project that John sent me. I want to implement it on our existing MacStories tools foundation. It needs to become its own separate tool based on the same stack as the other ones that we have, the images, the press release tool, the iOS review search tool, all of those. And, um, you need to just, uh, and I give it some boundaries, like you need support multiple users.

Each user needs to have its own database. It's a, each user needs to, needs to be able to store securely their GitHub credentials, like all of that. And the, so long story short, we do have editor as a, as a new tool in our arsenal of internal tools. The biggest challenge for Fable was figuring out the complicated history of documents when they are synced to GitHub. So the whole like commit push and like keeping track

**John** [16:14]: to

**John** [16:53]: of ownership,

**Ticci** [16:56]: especially because every once in a while we go into our repo for club MacStories, for example, and on a rolling basis, I think on a weekly basis, actually.

**John** [17:06]: Yeah. As soon as it's published.

**Ticci** [17:07]: As soon as it's published, we take the documents that belong to say weekly 520. And we move all of those into an archive subfolder. So at that moment, the path of the document actually changes. so keeping track

**John** [17:22]: And of

**Ticci** [17:23]: this document was created by Devon, but then at some point, John moved it into the archive subfolder. If you only consider the archive subfolder, it looks like the creator of the document was John, but actually it wasn't John. It was Devon because that document used to belong to another folder and it was moved. So all of those things can get quite nasty on GitHub, but Fable figured it all out. And of course, a lot of input from me was necessary to get right.

**John** [17:52]: this Yeah. And

**Ticci** [17:54]: the, yeah. So those were the projects that I created. One for MacStories, which we can now use, and it's working remarkably well, I think. And, uh, one that will eventually become a product later this year that I hope to release to the public and, uh, Fable just made it a lot, a lot nicer.

**John** [18:16]: Nice. Nice. The other thing I did was I also tuned up my, uh, macOS review research project a little bit because I had had a couple of problems with it. And, you know, none of these things really is ever fully finished because after coming home from WWDC, I thought of some other things I wanted it to be able to do. And once I started using it more intensely, I found a few bugs and things.

And I had it, I had a Fable five go through my Obsidian vault and just kind of audit where everything was and what things had been properly handled by the scripts that I have running on my Mac mini and which ones hadn't had it go and fix a couple of broken scripts and then backfill all the metadata and everything that had been missed and, and tuned it up that way, which was really good. I think too, it's worth mentioning before we move on Federico is that you and I use very differently, which I think is kind of interesting. I think one of the reasons you prefer codecs so much more than, than Claude, which is that I use Claude almost exclusively in the terminal. I do not.

I mean, and I think that that's because the Claude app just isn't great. And that's something that I think we can

**Ticci** [19:31]: talk about. Yeah. It's more of a, it's more of a statement on my end. I refuse to use a terminal if a GUI option is available. And if the GUI option is terrible, I'm going to complain about it

**John** [19:41]: until

**John** [19:44]: I get it. And I, and I, you know what, I think I, it was a breath of fresh air using Codex compared to Claude Code, which I did use from time to time.

**Ticci** [19:52]: It's really, really bad.

**John** [19:54]: Yeah. But I don't know what it was that drove me to use the terminal more, but I just, I always felt more comfortable using Claude in the terminal. And it was probably just because I didn't really like the app all that much. And, and I stuck with it, but that's one reason I think I stuck with it longer than you did because you had that friction of constantly going in that app and not being for you.

So.

**Ticci** [20:17]: Yeah. Even the idea that, I mean, besides the fact that the Claude for Mac app is so buggy, they've made it a bit better, but it's still very buggy. Like the fact that it doesn't, um, it doesn't always keep a reliable, uh, remote control connection. Like for example, when Fable was done redesigning Mac remote,

**John** [20:39]: I

**Ticci** [20:39]: was managing that session. I created a session on the Mac because another dumb decision, you cannot create a new Claude Code session in a folder on your computer from your iPhone. You need to create it on the Mac and then remote control is enabled by default. The issue is that it's so buggy. I was managing the session from my phone in bed the night before. went

**John** [21:04]: I to

**Ticci** [21:05]: sleep. I woke up, I opened Claude on my phone. The session is not there. I VNC into my Mac Studio. I see that Mac remote had been rebuilt and redesigned. So I opened the Claude app on the Mac and I realized for some reason, the session had been archived

**John** [21:21]: and

**Ticci** [21:22]: therefore remote control had been disabled. I had to unarchive the session, send a new message from the Mac app, and then remote control became available again. So that kind of stuff is just so bad. Yeah.

**John** [21:37]: I think one of the things too, that was this past week was announced by Anthropic was this new thing where they call it like reflections, where you get some statistics on how you've been using Claude in the Claude app. And I saw it and I was like,

**Ticci** [21:50]: To make you feel bad about the usage?

**John** [21:53]: I don't know. Well, I saw it and I thought, you know what, this is kind of interesting, but couldn't you just like fix some of the stuff in the app first before you do what is essentially like, you know, bragging rights about how you use Claude or something? I don't know. It is kind of too bad.

**Ticci** [22:12]: Yeah. They have interesting priorities at Anthropic for sure.

**Ticci** [22:19]: This episode of AppStories is brought to you by Daylite. If you run a business on a Mac, you know the challenge. Client emails are in one place, notes are in something else, and projects live in another place altogether. And when something falls through the cracks, it's usually because it's in your head or scattered across post-it notes instead of in your system. That's exactly what Daylite was built to solve. For over 20 years, Daylite has helped service-based businesses keep up with their clients, projects, tasks, calendars, and communications all connected in one place.

Every email, meeting, note, file, and follow-up becomes part of a complete and ever-growing client history, giving you the context you need to know what happened, what's next, and who needs your attention instantly. And with Mail in Daylite, your email lives inside the rest of your work, so conversations naturally stay connected to the people, deals, and projects that matter. Built exclusively for Mac, iPhone, and iPad, Daylite feels right at home in the Apple ecosystem. If you're ready to spend less time searching for information and more time moving your business.

Visit daylight .app and start your free trial and use the code APPSTORIES50, that's APPSTORIES50, to get 50 % off your first year of Daylite on any annual plan. That's D-A-Y-L-I-T-E .app and the code APPSTORIES50 for 50 % off your first year of Daylite. Our thanks to Daylite for their support of the show.

**Ticci** [24:09]: So, the other things that I built, I built with Fable in Cursor.

**John** [24:16]: Yeah, hear about that.

**Ticci** [24:16]: we'll So, I was fed up with the limits. I spent a lot of credits on extra Fable usage in Claude.

**John** [24:27]: Especially with the extra usage.

**Ticci** [24:29]: Oh my God. It

**John** [24:29]: burns those up fast. Because it's basically API rates, which much higher.

**Ticci** [24:34]: are Yeah. And I was fed up with that. And I was annoyed with the Claude for iOS experience for remote control.

**John** [24:48]: So,

**Ticci** [24:49]: I figured I'm just going to give Cursor a go. Cursor, they have really generous usage tiers. And they let you use Fable. And I've long been curious about the agentic loop and the harness that Cursor uses. And I built a couple of things. I started a couple of things in Cursor. And I'm now finishing them with Sol in Codex.

So, the first one that I did, I created months ago with Codex, another MacStories internal tool that uses the complete archive of my iOS reviews in Markdown. To let me search for any detail from my iOS reviews. Any sentence. Any footnote. Anything that I can think of.

Using a vector database. It uses RAG to turn my archive of iOS reviews into an embedding. So, a vector database.

**John** [25:55]: Yep.

**Ticci** [25:56]: And it was using GPT-55 to let me ask the question in natural language. Like, for example, how many times... Typical question. How many times did Apple redesign Control Center over the years? And you wait a couple of seconds. And since it's based on a vector search, it's really fast. And so, in five seconds, you could see an answer saying, based on your reviews, Apple redesigned Control Center, I don't know, three times.

That

**John** [26:22]: would have been my guess, would be three. And

**Ticci** [26:24]: Three times. it would have, like, the snippets, the relevant snippets from the database comprised of my own reviews. I built this months ago in GPT-5 .2, maybe. And then refined it with 5 .5. Pushed it up to the MacStories Tools page that we have. And I've been using it a lot to get ready for iOS 27.

And it was working quite well. It was based on the OpenAI embedding model. What's it called? Embedding large V3. And GPT-5 .5 to ask questions. But that was it. Like, I could only ask a question and see the answers. And I wanted to have three things. One, I wanted to see if I could use a better state-of-the-art embedding model.

Two, I wanted to have integration with the web so that in addition to searching my local archive. So imagine that these are, like, 11 markdown files that have been turned into a vector database. But they don't contain the actual links back to MacStories, right? They're an offline archive. So if I'm asking a question about, like, when did Apple introduce digital touch in iMessage? It can't give me an answer. It cannot give me a source link to the exact page on MacStories .net where that sentence is mentioned, right? Because it's only searching an offline plaintext archive.

So I wanted to have a system where I could click an optional button to say, okay, now find this result on MacStories and the exact page and the section where it appears.

**John** [28:22]: Yeah, that's nice.

**Ticci** [28:24]: And third, I wanted this feature to have a fancy animation.

**John** [28:28]: Of course you did.

**Ticci** [28:30]: So I did some research, found out about this Voyage embeddings model. called? Voyage AI.

**John** [28:37]: What's it

**John** [28:39]: Voyage or Voyager, I forget which one.

**Ticci** [28:42]: Voyage, I think. Voyage AI. They have this Voyage 4 model.

**John** [28:43]: Voyage, yeah.

**John** [28:47]: Yeah.

**Ticci** [28:48]: So I'm using that. It's really nice. It's really good results. And I built the web integration by connecting to SERP API and Firecrawl so that after I click a button, GPT can use those tools to run multiple searches and find an exact quote from my review on the actual MacStories .net page.

**John** [29:15]: How come you're using both of those? Don't they do the same thing basically?

**Ticci** [29:20]: Like one is the fallback of the other, I think. it. It defaults to Firecrawl. Firecrawl cannot do it. It goes to SERP. If SERP cannot do it, it tries its own GPT web search thing.

**John** [29:23]: Got

**John** [29:32]: Okay.

**Ticci** [29:33]: So like it has options to check. When this checking is happening, I had Fable design a fancy WebGL animation.

**John** [29:34]: multiple

**John** [29:45]: So

**Ticci** [29:45]: WebGL is using like, it's like GPU-based acceleration for animations in web browsers. And so when I click this button, I see this animation that is like a network. It looks like a neural network. With all these dozens of lines moving. And as they move, as the model progresses, it's actually synced in real time. So as the model nails down the page

**John** [30:01]: Okay.

**John** [30:12]: where

**Ticci** [30:13]: the result appears, the lines condense. that's cool. Until it just becomes one line. And then the result pops on screen.

**John** [30:17]: Oh,

**John** [30:21]: That's cool.

**Ticci** [30:22]: So that was really nicely done. I love it. Again, Fable. All of this was done in Cursor. And I was really impressed by the fact that Cursor, you can do things like, I would prompt stuff like, Fable, you are the orchestrator. I want you to spawn, like, I don't know, a team of Composer. Composer is the in-house Cursor model. Like, I want you to spawn a team of Composer sub-agents and you're going to review their work in a loop.

So you're going to make a plan. Fable makes a plan, hands it off to Composer sub-agents. Each of them gets to work, goes back Fable. Fable evaluates the work and says, oh, this is not good enough. Go back. And it gives, like, a list of fixes. And that sort of sums up my experience with Fable. I think Fable is obviously great at writing code, but it's potentially even better at just orchestrating smaller and therefore cheaper models.

Making plans and acting as a coordinator for those agents.

**John** [30:57]: to

**John** [31:20]: Yeah. If you use it as the only tool that you're using to code, you'll end up in the poorhouse pretty fast. I mean, you really need to let it. And it's actually designed in a way that it's very good at deciding which model it should use in terms of, like, the, you know, the quality of the model, but also the thinking level and all other parameters. So that, I found that to be really good too. Because I did do the thing at the beginning where I was just like, oh, Fable, do this whole thing for me. And it's

**Ticci** [31:40]: those

**Ticci** [31:48]: like, well. No, that's not a good

**John** [31:49]: idea. That's not a good idea.

**Ticci** [31:51]: You can do that with Sol. I do not recommend doing it with Fable. The other thing that I will quickly mention, this is just a small teaser. Okay. I'm building an Apple Notes CLI. So the CTL family of Federico CLIs is growing. After

**John** [31:55]: Yeah.

**John** [32:08]: REST

**Ticci** [32:09]: CTL, I will be doing Notes CTL in terms of Apple-specific command line interfaces. started out with Fable. I am continuing with Sol. We'll talk about that in the future. The final thing I will say about Fable before we talk about OpenAI is that you can tell that it's a different beast. And I've been thinking about this. Like, how can

**John** [32:18]: I

**John** [32:37]: you tell?

**Ticci** [32:39]: Even if the benchmarks are comparable between Fable and Sol, I still think Fable has a different feel. And by that, I mean, it's obviously a bigger model, right? Some people are saying it's a 5 trillion parameter model. Other folks are saying, no, it's a 10 trillion model. It's obviously bigger than GPT-5-6 Sol.

Sol is not based on a new, larger, pre-trained foundation from OpenAI.

**John** [32:57]: Wow.

**John** [33:08]: It's more of an evolution of what was... It's

**Ticci** [33:10]: an evolution of 5 .5. Rumors are saying that that bigger model from OpenAI will be GPT-6, potentially even by September. So we'll see. But Fable is a different beast in the sense that I think here's sort of how I've sort of started understanding these things. You can get a feel for the size of the model by its quote-unquote intuition. Like how it tries to answer the unsaid in your

**John** [33:50]: request. Right. Understand what you mean as opposed to explicitly said.

**Ticci** [33:53]: what you Exactly. It has a bigger sense of trying to guess your underlying intent. Even if you don't specifically say, I want you to do this and this. For example, when I asked Fable to redesign Mac Remote, I said absolutely nothing about researching the optimal design for an onboarding flow.

**John** [34:19]: Right.

**Ticci** [34:20]: I said nothing about tutorials, about splash screens. It just guessed that that was part of my request. Whereas even with Sol, you will see that you need to hold its hand a little bit more.

**John** [34:30]: Right.

**John** [34:39]: Yeah. Yeah. Yes. Yes. More complete instructions.

**Ticci** [34:41]: Right.

**Ticci** [34:44]: Yes. When you hear people say, oh, it's got the big model smell, usually that means since it contains more knowledge, right? Because it all comes down to that. Like when you hear trillions of parameters, it means it was trained on more stuff. contains more knowledge inside. And typically that means a model is able to have an intuition for your goal, even if you don't exactly specify the multiple steps of your goal.

**John** [35:00]: Yeah. It

**John** [35:18]: Right.

**Ticci** [35:18]: So that was my experience with Fable.

**John** [35:21]: Right. It's an approximation of experience is the way I kind of think of it.

**Ticci** [35:24]: Yes. You know

**John** [35:25]: what I mean? In a certain way.

**Ticci** [35:26]: Like when you talk to, if I was assigning this project to an experienced software engineer and designer.

**John** [35:36]: You wouldn't have to tell them to do a splash screen, right? the thing.

**Ticci** [35:38]: That's The person would have said, yeah, obviously part of that, right? And that sense of taking things for granted or intuition, whatever

**John** [35:41]: it's

**John** [35:50]: you want to

**Ticci** [35:50]: call it, that is a distinctive trait of a bigger model.

**John** [35:51]: usually

**John** [35:56]: Right.

**Ticci** [35:56]: That's how I would

**John** [35:56]: put it.

**Ticci** [35:58]: Now, do you want to talk about OpenAI?

**John** [36:00]: Yeah, we should. We should talk about the models. And I do want to briefly touch on the app too, because. let's

**Ticci** [36:05]: Yeah, start with

**John** [36:06]: that. You want to start with that?

**Ticci** [36:08]: Yeah. Yeah.

**John** [36:09]: Let me set the stage for You know, OpenAI has said a lot of things over the months. They've said that they wanted to build an OS. They've said that they wanted to build a super app. And then Codex came out, and we all sort of assumed that that was the super app. And it kind of was. But Codex no longer exists, because what OpenAI did when they released these new models is they took Codex.

They merged it in with ChatGPT. They merged it in with ChatGPT. But what was left after the merger was more the bones of Codex. Because now you can toggle between two things. You can toggle between ChatGPT Work, which is a new feature, which is kind of like Claude Cowork if you've ever used it. The idea is taking some of the agentic coding tools, applying them to productivity tasks, and creating a little bit more of a protective harness around them that is not

**Ticci** [36:11]: this.

**Ticci** [37:11]: going

**John** [37:11]: to get people in trouble as easily as if they were using Codex to do spreadsheets. So you can toggle between those two things. Codex, which already existed for writing code. ChatGPT Work for doing productivity tasks. And then they have in the sidebar, it still says ChatGPT. And you think, okay, well, what's that? And how is that different than ChatGPT Work? Well, ChatGPT is just ChatGPT. It's the chatbot that's been around for quite a long time now. And when you click on that in the sidebar, what happens is that a little window, on the Mac at least, pops up in the corner and you can do your chats over there.

And I think I understand where they were coming from with this design. I think it's a bad idea. But I think what you have is a bunch of engineers who probably are working in Codex all day and use ChatGPT for little side research things every now and then. And so having it pop up in the corner, answer a quick question, then go back to Codex and your coding work, I think that makes a lot of sense for a person like that.

They aren't using ChatGPT that way. Most people are like, tell me when the restaurant down the street opens on Saturday, you know, or something like that. It's not that kind of thing. And so you can pop it. It can like kind of hover in the corner or it can pop up and be more of its own window. And it's, I don't even know if you can change the model there. It's definitely, it's tuned to be fast, which is good.

But people are pretty upset about it. And I know that you've run into quite a number of bugs beyond just kind of the design decisions that have been made here.

**Ticci** [38:54]: Yeah. So first of all, I recognize that it's a really tough challenge for any designer right now to make sense of this effectively three modes that we have. Because if you look at OpenAI and Anthropic, they are facing the exact same issues. We are moving into this direction with these two companies where the models can be accessed in three different ways. You have the regular chatbot experience with memories, projects, all of that stuff. You have the coding agents.

So Claude Code and Codex based on a different harness, right? It's all about the harness. It's all about like the harness is the structure around the model that controls how it does multiple things like an agent in the background, multiple steps, right?

**John** [39:50]: Tools that can and how they can

**Ticci** [39:51]: access access Which is not exactly what a chatbot does. And they're based on a different system prompt. Like they're very different. But now there's this third category, which is Claude Cowork and ChatGPT Work, which is inspired by the coding agent harness without the complexities of programming and code for developers. So both companies sort of arrived at the same conclusion of like, well, this agentic loop is really good, but it's too confusing for knowledge workers or just people who want to use it for productivity tasks instead of programming tasks. How do we take that and make it about general work?

So now you have these three surfaces, the chatbot, the coding agent, and the work agent. How do you make sense of those?

**John** [39:53]: them.

**John** [40:46]: This is the classic everything app problem that we many times. Go

**Ticci** [40:49]: talked about figure. Yes. So I don't have a good answer.

**John** [40:56]: No.

**Ticci** [40:56]: But I can tell you this. I think we are now in a really curious scenario where right now, as of Friday, July 10th,

**Ticci** [41:10]: and I didn't think I would ever utter this sentence on AppStories. I think right now, 4 PM, July 10th, Anthropic has a less confusing design for all of this than OpenAI.

**John** [41:25]: That said, they are both confusing in own ways. Look,

**Ticci** [41:28]: their they're both terrible.

**John** [41:33]: I hear you. No, you're

**Ticci** [41:34]: right. If you ask me at gun points, John, one do you prefer? You've got to pick one. I

**John** [41:37]: which

**John** [41:40]: would answer Not

**Ticci** [41:41]: Claude. the model, just the UI of it all.

**John** [41:44]: You might wonder, why is OpenAI doing this? Why are they throwing their bread and butter under the bus by demoting it to the corner of the Mac app? And I think it's pretty clear. I mean, they have 900 million users of ChatGPT. Codex has 5 million. And they all right?

**Ticci** [42:03]: want, You know,

**John** [42:05]: 5 million is a

**Ticci** [42:06]: lot of people. They want money.

**John** [42:07]: It's a lot of people. But that's where they're making money. And ChatGPT, not so much. So what they want to do is expose Codex to the ChatGPT people. I just think that this is kind of a ham-fisted way of doing it.

**Ticci** [42:20]: And there's a little guy called Siri AI on the horizon that will potentially become a pretty decent, normal chatbot. Of those 900 million people using ChatGPT, let's assume, and I'm being overly optimistic, let's assume that 500 million are using it for free. It's probably more.

**John** [42:46]: Oh, I bet more, a lot more.

**Ticci** [42:47]: it's But let's say that half a billion people are using it for free.

**John** [42:52]: Yeah.

**Ticci** [42:53]: And let's face it, they're probably not the kind of nerds that listen to AppStory. So I can for sure we don't have half a billion listeners. Their usage of AI is probably quite average and simple, which is not meant in a derogatory way.

**John** [42:58]: tell you

**John** [43:10]: It's

**Ticci** [43:10]: totally fine. And for those people, Siri AI can potentially be good enough. better. And or better.

**John** [43:16]: Or

**John** [43:19]: Because of the integration with apps.

**Ticci** [43:22]: And if I were OpenAI, I would be seriously concerned about that. Therefore, I would say, well, we got to make money.

**John** [43:30]: Right. Well, they look at it and they say not only, and it's like a Google thing too, because, you know, Google has Gemini on Android. And now they're working with Apple in a lot of ways. You could argue, as I think, you know, Chris Lawley, we were talking to him yesterday. He made the point.

Google's locking up AI on mobile. And, you know, I guess OpenAI sees the running on the wall potentially here.

**Ticci** [43:31]: and

**Ticci** [43:53]: Potentially, yeah. So, but I think they kind of fumbled it. So, the main issue is that everything has been renamed to ChatGPT. Right. Makes sense. It's the name with the most weight of all of them. And you're doing ChatGPT and ChatGPT Work, but you're keeping codecs because it's already an established name in the programming scene.

Right off the bat, the first problem is that you renamed codecs to ChatGPT. You updated ChatGPT to ChatGPT Classic. You're redirecting people to the new ChatGPT app. But when they open it, all of their chats and folders of the projects of chats, they're gone.

So, all of the things that is, if you want to have a regular ChatGPT experience right now, you have to go to ChatGPT .com in a browser. There, you will still see your chats and your projects.

**John** [44:04]: Right.

**John** [44:56]: It's better on the iPhone too. It's better on mobile. It's easier

**Ticci** [44:59]: to get to them. Well, they kind of fumbled it on the Mac. Right now, the Mac, they really shove it in your face. Like, they really want to show you, oh, now you can work with the file system.

**John** [45:10]: Because that's what you do. You get real work done on a Mac, Federico. I mean, all.

**Ticci** [45:13]: after Sure. They got rid, did you see, John, they got rid also of the floating window thing for ChatGPT for Mac. Like, you would press a hotkey on your Mac and you could ask a question to ChatGPT.

**John** [45:28]: Oh, they got rid

**Ticci** [45:28]: of that? After that, got rid of the feature to record meetings and transcribe them.

**John** [45:30]: they

**John** [45:35]: Wow.

**Ticci** [45:36]: Yeah. make

**John** [45:37]: That doesn't sense.

**Ticci** [45:38]: Yeah. And obviously, they went from a native Mac app to an Electron app, which doesn't bother, I think, us as much as it does other people. But something worth keeping in mind.

**John** [45:48]: it's It's a much bigger app if about

**Ticci** [45:52]: you care the size. 1 .5 gigabyte download also. But the main issue is that, like Anthropic, they are now facing this challenge of how do you reconcile these three modes that, on Mac, web, and iOS, on each platform. So there are three modes. Think about it this way. There

**John** [46:18]: are three modes. Even within modes, there are differences. hold on.

**Ticci** [46:21]: Hold on, There are three modes.

**John** [46:24]: Yeah.

**Ticci** [46:25]: With three models each, right? Yeah. Luna, Terra, and Sol.

**John** [46:30]: And then probably like five different thinking levels.

**Ticci** [46:32]: Five thinking selectors. And on each platform, again, Mac, web, and iOS, they have different limitations. They For example, ChatGPT apps, you may think, wait, well, ChatGPT used to be called MCP, right?

**John** [46:45]: do.

**John** [46:56]: Yeah.

**Ticci** [46:56]: Then they were called connectors. Then OpenAI last Christmas, they did a whole thing, ChatGPT apps. They have been renamed to plugins, which again, makes sense, right? They were called plugins and Codex. Now they're called plugins everywhere. Cool. But the plugins that you enable on the web, they don't appear on the Mac.

**John** [46:58]: Yeah.

**John** [47:20]: Yeah.

**Ticci** [47:21]: Because they're using different sandboxes.

**John** [47:23]: Right. Well, that's the same problem they have with ChatGPT Work too.

**John** [47:32]: It's sandboxed in one way, but if you use it on the Mac, it's sandboxed another way. So you won't see the stuff you do on your Mac on mobile or on the web, as I understand it, and vice versa.

**Ticci** [47:45]: Yeah. The memories that you have in ChatGPT .com. I'm just going to use ChatGPT .com as a way to refer to the regular chatbot experience. Your memories that you've probably accumulated over the past three years, they do not carry over.

**John** [48:03]: To anything else, right?

**Ticci** [48:04]: To anything else.

**John** [48:05]: Yeah.

**Ticci** [48:06]: On iOS and iPadOS, the codecs tab has been renamed to remote. Right?

**John** [48:16]: Right.

**Ticci** [48:17]: So on iOS, I actually think on iOS it makes more sense. Because on iOS, now you create, well, it kind of makes more sense.

**John** [48:24]: But you can't remote into anything but codecs, right? I mean, you can't remote.

**Ticci** [48:28]: Exactly.

**Ticci** [48:31]: This is why I think it's kind of better, but not totally. So on iOS, now you open ChatGPT, you can either chat or work.

**John** [48:36]: Yeah.

**John** [48:41]: Yeah.

**Ticci** [48:42]: Work is the cloud sandbox.

**Ticci** [48:48]: Which is pretty

**John** [48:49]: limited.

**Ticci** [48:50]: Yeah. The other thing that I don't understand right now in ChatGPT for iPhone, when you go to, at least on my end, maybe I'm being A-B tested. I'm not sure. When I create a chat, I cannot choose the model anymore. In regular chat mode,

**John** [49:02]: Really?

**John** [49:04]: I

**Ticci** [49:05]: swear, I do not see a model picker. The only way that I can have a chat about chat right now is if I go to the website. Interesting. In ChatGPT for iPhone, I don't see a model picker.

**John** [49:08]: Hmm.

**John** [49:18]: No, I don't see a model picker. Yeah, there's no model picker on the Mac either. I'm looking at it right now.

**Ticci** [49:25]: Yeah. It's

**John** [49:26]: just, it's called Instant. Oh no, I do have a picker. I have a picker. Well, no, it's, hmm. Yes, I do have a picker. It's just different.

**Ticci** [49:33]: Okay. On the iPhone and the iPad, when I go to remote, that lets me remotely connect to my machines, my Mac and my MacBook Pro.

**John** [49:41]: Studio Right.

**Ticci** [49:44]: But that is only for Codex threads.

**John** [49:50]: So,

**Ticci** [49:51]: look.

**John** [49:52]: It's too much to people to keep in their head.

**Ticci** [49:55]: This is exactly what was going to happen. This is exactly why we did an episode about the Everything app. We knew this was going to happen. It gets messy fast. It gets messy fast. Here's my proposal. Keep the stuff for programmers in its own separate thing.

There's probably no great way to reconcile the needs of developers and tinkerers and programmers with the much bigger ecosystem for regular productivity work, knowledge workers, office workers. I understand the sentiment behind blending, unifying ChatGPT and ChatGPT Work. Codex shouldn't be part of it. I think if you ask me, that's the mistake. Just like for Anthropic, I could totally see Cloud being chat and cowork, but not being Claude Code in the same app. I think these coding agents, they need to be their separate environment.

**John** [50:00]: Yeah.

**John** [51:00]: Right.

**Ticci** [51:00]: And you make peace with the idea that you're going to have to maintain two separate applications. This is way too confusing.

**John** [51:08]: Yeah. I mean, you know, it's like, I think it's like almost a historical accident that we are where we are because, you know, the coding agents came first. So they were dumped into the apps first and there is a lot of crossover between the functionality of what you would do with cowork and Claude Code, for instance, because, you know, they were inspired. You know, the cowork was inspired by Claude Code, but it doesn't mean it has to be in the same app.

**Ticci** [51:33]: Yeah. Yeah. And we're bending over backwards to try and make sense of these different modes and different models and Mac web and iOS where things could just be a whole lot simpler if there were separate apps. Yeah. Anyway, do you want to talk about the models in AppStories Plus? Maybe we can save it for

**John** [51:48]: Yeah.

**John** [51:53]: our listeners? Yeah, let's do that. Let's do that. All right. Well, thanks again, everybody, for joining us today. Another great episode of AppStories. And thanks, too, to our sponsor, Daylite. Federico over at MacStories .net. And, of course, we're on social media where Federico is @viticci. That's V-I-T-I-C-C-I. And I'm @johnvoorhees. J-O-H-N-V-O-O-R-H-E-E-S. Talk to you next week, Federico.

**Ticci** [52:19]: Ciao, John.