This Week's Sponsor:

SaneBox

Clear the Clutter, Keep What Matters.

Building for iPhone Duo, in App Developers’ Own Words

I spent yesterday in New York City interviewing developers about their iPhone Duo apps and the experience of building for such a new and different device. The depth and breadth of developers Apple assembled was excellent. I spoke to indie developers like Vidit Bhargava about Deck-Man and the team behind Notewise. I also interviewed developers from big companies about their impressive Duo updates, including The New York Times for its NYT Cooking app, Netflix, and Meta’s Edits video editing app. Plus, I met Mark Otero from Azra Games, whose team poured their hearts into UnGodly, one of the best-looking games I’ve seen on iOS in a while.

A few things struck me about my conversations with developers. First, every one of them was excited about the design challenges posed by the iPhone Duo and the many orientations and ways it can be used. Second, every one of them came away impressed by how smooth Apple has made the development process. That’s not to say the design challenges weren’t real, but getting from a design idea to an app that worked on the Duo went better than anyone expected. Third, after months of building and testing, everyone was looking forward to seeing which design choices land with users.

But that’s enough of my impressions. I’ll let the developers speak for themselves in this special bonus AppStories video. I came away from New York with some great interviews, which I’ve shared on the MacStories YouTube channel. Apologies for the sometimes rough audio. There’s only so much I can do with recordings made in a room packed with press and developers, but if you’re into apps, design, and how things are built, these interviews are for you.

Also, I left New York with a Night Sky 512GB iPhone Duo that I’ve only just set up and begun testing. I’ll be writing about the Duo soon, and we’ll start covering the best of the more than 300 Duo-optimized apps that developers have sent Federico and me. Stay tuned, because the busy fall season at MacStories is just going to keep rolling.

iPhone Duo Review Roundup: Great Hinge, Messy Apps, and Everybody Misses Face ID

iPhone Duo reviews are out today, and while I was hoping that this year I’d get to review this hybrid device under embargo, it hasn’t happened (I hope it will, someday!). I’ve spent the past few hours reading, annotating, and watching as many Duo reviews as I could, and I’ve compiled the most interesting bits and common themes in the roundup below.

The broad strokes are consistent across all the reviews: the Duo’s hinge and the matte inner display are terrific; the lack of Face ID and a telephoto camera are serious trade-offs; and most third-party apps aren’t ready yet. But there’s plenty of nuance in the details, and even some disagreement (about the speakers, of all things).

Let’s dive in.

Read more

Perplexity Computer’s Local-First Approach to Agents

Source: Perplexity.

Source: Perplexity.

Today, Perplexity announced a local-first mode for Computer where the model and agent work entirely on your Mac whether you’re online or not, sending data to the cloud only with the user’s approval. The new local mode complements Hybrid Compute, which was introduced last month and uses a combination of local models and cloud-based frontier models.

The result is that Perplexity customers now have a choice when they start up the company’s Mac app. They can choose to start agentic workflows in the cloud with frontier models, or they can flip things around and start locally, turning to cloud-based models as needed. The local model uses Qwen 3.6 or PPLX Pro, a post-trained version of Qwen.

The benefit of starting locally in a securely sandboxed code- and tool-use environment is twofold. First, data stays on your Mac, ensuring privacy. Second, Perplexity subscribers are able to conserve credits, making their subscriptions go further. At the same time, cloud-based models remain available with the user’s permission as needed and to perform tasks like browsing the web and accessing online apps.

Local Computer is available to Perplexity Pro, Max, and Enterprise subscribers now, and Perplexity recommends using Apple silicon Macs with at least 36GB of memory.

Apple Announces October 28 “Pro and Tell” Event

In a post by Greg Joswiak on X, Apple has said it will hold an ‘Pro and Tell’ in Los Angeles, California on Wednesday, October 28 at 10:00 am Pacific.

Judging from the ‘Pro’ in the tagline, I wouldn’t be surprised if we see the much rumored touch-screen MacBook Pro at the event. Other rumors point to a possible refresh of the iMac and the iPad mini.

Our very own Federico will be there in person to cover the event, and of course we’ll have written coverage here on MacStories and later on AppStories.

ChatGPT Dots Gain Widget and New Shortcuts Actions

Late Friday, OpenAI updated the ChatGPT app on iOS and iPadOS with a new widget for your Dot and Shortcuts actions that let you control it three ways. The implementation is really well done and adds automation options worth considering.

First of all, the Dot widget is about as simple as it gets, but also very useful. I have the ChatGPT app set up to drop me into its Codex view because that’s typically how I use the app. I open Codex, check on a long-running project, and provide whatever input is necessary to keep things moving.

The problem I’ve run into is that I’ve been using my Dot, Puck, more and more to kick off research, ask one-off questions and check in on things like my email. Landing in Codex means tapping around more, which I can avoid now with ChatGPT’s new small-sized widget. The widget has two modes. The first simply shows your Dot’s avatar, and tapping on it opens your Dot’s message thread. However, if you edit the Dot, you can toggle on the “Show shortcut button” to display messaging and voice buttons beneath your Dot, adding the option to start a voice call with your Dot. As simple as this is, it’s been great to have the option to start in Codex via the ChatGPT app or check in with my Dot thread using the widget.

The other interesting feature is a trio of Shortcuts actions to message, dictate, or call your Dot. Each action has two parameters: the name of your Dot and how you want to communicate with it. That first parameter is an interesting one because tapping on it opens a list with your sole Dot in it because currently, you can only have one Dot. However, it also strongly suggests that the ability to set up multiple Dots is on the horizon, which OpenAI clearly telegraphed during its DevDay. After all, why have a parameter that opens a list unless there’s more than one choice?

I've set my Action button up to call my Dot.

I’ve set my Action button up to call my Dot.

The actions can be used in your shortcuts, are accessible from Spotlight Search, and most interestingly can be tied to an iPhone’s Action button, allowing you to visit your Dot’s thread in any of three ways. I’ve set my Action button up to start a Dot call. I often think of tasks and have ideas for research topics when I’m out for a walk, and I want to see if having brief calls with my Dot is an effective way to get a jump on those ideas, so they’re waiting for me in ChatGPT when I get home.

These actions and widgets are an excellent addition to OpenAI’s app and a cut above Anthropic’s more barebones approach to widgets and Shortcuts. I would, however, like to see Dots integrated even deeper into Shortcuts. The Shortcuts actions only open your Dot thread in one of the three modes. I’d love to be able to connect Dots to other apps with actions that accept input from other sources and produce outputs that can be used in other actions.

While I’m making requests, I still think multiple Dots need to be supported sooner rather than later, and OpenAI needs to provide a better mechanism for switching between Macs if you use a Dot with multiple computers. Plus, Dots are still a little buggy, with scheduled tasks failing too often and for reasons that aren’t always clear.

That said, the latest update is worth checking out if you’ve found Dots helpful. The strength of Dots is that they’re everywhere you can access ChatGPT, and now with widgets and Shortcuts actions, getting to your Dot in each of those places is easier than ever.

SaneBox: Clear the Clutter, Keep What Matters [Sponsor]

You’ve set up Focus modes for work and weekends, built Shortcuts for repetitive tasks, and rearranged your Home Screen until everything is perfect. But then your phone buzzes with another notification for a newsletter, shipping update, or sale. Meanwhile, the important email you need to follow up on is buried under everything else. You’ve put thought into how your devices work for you, but are you in charge of your email or is it in charge of you?

SaneBox brings that same attention to detail to your email. Its AI filtering sorts your messages with nothing to install, and since 2010, it has helped people separate the messages that need attention from the ones that can wait.

SaneBox’s SaneLater folder keeps messages that can wait in a dedicated folder, so you can focus on more important things. SaneNews collects newsletters in one place, ready for when you can take a break. And if the same unwanted sender keeps interrupting your day, move one of their messages to SaneBlackHole to keep their future messages out of your inbox.

There are tools for staying on top of ongoing conversations, too. SaneNoReplies collects conversations where you haven’t received a response, so you can follow up. When you need an answer before a deadline, SaneReminders can remind you to check back. You can snooze messages, too, putting something aside until you have time.

SaneBox works with the account and apps you already use. It sorts messages on your mail server, so its folders are available in Apple Mail, Outlook, or whatever other email app you use on your Mac, iPhone, and iPad.

SaneBox also learns from your email habits, and you can teach it your preferences. If a message belongs somewhere else, move it and SaneBox will use that folder going forward. Plus, the Daily Digest brings together unread messages in your SaneBox folders, so you can review everything when you’re ready to wrap up your day.

Privacy matters, too. SaneBox sorts using message headers, not the contents of your emails, and it doesn’t store full messages or attachments. Its security is reviewed every year as part of Google’s verification process, audited independently by Leviathan Security Group, and tested through HackerOne.

All of that adds up to less time managing email and more time for everything else, which is why more than 100,000 people rely on SaneBox and have given it a 4.9 rating across more than 1,000 reviews.

Best of all, SaneBox fits elegantly into the email system you already know and use. Start a seven-day free trial with no credit card required, and get a $25 credit toward a SaneBox subscription, the equivalent of three months of use.

Our thanks to SaneBox for sponsoring MacStories this week.

Meta Releases Muse Spark Open, an Over 800 Billion Parameter Model That Can Run Locally on Multiple M5 Ultra Mac Studios

Meta just released the weights for Muse Spark Open, a multimodal large language model with a mixture-of-experts architecture. It has over 800 billion parameters in total, with 40 billion active per token, and it can run on the latest Mac Studios.

Meta worked closely with Apple to make that happen specifically on the M5 Ultra Mac Studio with the support of MLX, the Apple-led, open-source machine-learning framework. The result is a frontier-class model that can be run locally on desktop hardware.

Apple’s Machine Learning Research blog put the M5 Ultra through its paces standalone and connected with other M5 Ultra Studios, and the results are impressive. If you’re at all interested in running LLMs locally, it’s worth checking out all the details, but a few things jumped out at me:

  • The M5 Ultra’s time to first token exceeds the M3 Ultra by 4.6x thanks to the M5 Ultra’s Neural Accelerators.
  • Tokens generated by the M5 Ultra beat the M3 Ultra by 1.6x.

However, thanks to MLX’s support of RDMA over Thunderbolt 5, big models like Muse Spark Open can be split across multiple M5 Ultra Mac Studios accelerating tasks further. For very large contexts that means a time to first token that’s up to 1.8x faster, but it’s also what lets the model fit onto desktop hardware because even the 4-bit version of Muse Spark Open requires 493GB of memory.

Some of Apple’s testing was done with the 512GB model of the M5 Ultra Mac Studio, which is slated for release at the end of the year. Apple hasn’t said how much the new model will cost, but for context, the 256GB model with 2TB of storage will set you back over $11,000. So the 512GB model isn’t your typical consumer purchase, even more so since running Muse Spark Open requires at least two of them.

However, that doesn’t make this any less of an accomplishment. Muse Spark Open is at the bleeding edge of models that can be run locally on computers that can sit on your desk and don’t need more power than can be delivered from a wall outlet. That’s made possible by the unique combination of Apple silicon and updates to MLX. Plus, with the pace of advancements in AI and Apple silicon hardware, what may only meet enterprise budgets today could become accessible to more consumers faster than most of us might expect.

Why Aren’t AI Labs’ Apps Better?→

Ben Davis (co-host of one of my favorite new podcasts this year, Nerd Snipe) has a fascinating take on a question I’ve seen many in the Apple community ponder as well: if these frontier models are so good, why do most AI apps from the big labs still kind of suck?

Davis wrote about this in relation to the ChatGPT/Codex desktop app, which has turned into a bit of a kitchen-sink design phenomenon with all kinds of features crammed into a UI that has two left sidebars, multiple modalities, and all sorts of disjointed menus.

To me the most interesting thing here is the implied question of “if the Codex team has unlimited tokens on really powerful unreleased models, why is the product getting worse?”

The lazy answer is “it’s just vibe coded slop, real artisanal code wouldn’t be like this”. Anyone who still says this is a moron and you can ignore them.

“Vibe coding” is not the problem. Frankly, these models are better at writing code than pretty much anyone at this point. You tell the model to make something and it’ll make that thing quite well. And also to be clear, Jira was horrid before LLMs existed. Good and bad products have always existed, and I actually think you can make infinitely better products now when using the agents properly (see T3 Code and Grok Bot).

I’ve been reviewing apps for over 17 years. Without a doubt, the most basic “vibe-coded” app created with Opus 5.5 today is infinitely more stable and nicer than any first app from an amateur developer in 2009. So if we agree that these models can write good code now, why is it that so many apps are still bad?

As we argued in our interview with OpenAI’s own Andrew Ambrosino (the irony is not lost on me), it comes down to human discernment, taste, and other qualities that can’t be turned into “metrics” for a model…at least not yet.

Davis continues:

To put a more precise definition on it, I think this “taste” is the ability to make good decisions in places where there isn’t a programmatic verification function.

For example: “make this function really fast” isn’t a “taste” task it’s a hill climbing task (make changes => test changes => repeat until goal met) which the models excel at right now (I think that any problem which can be expressed in this shape can be solved now by Opus 5.5)

“Which of these color palettes do you prefer?” is a “taste” question. You can’t really put a verifier on it other than your own opinion AKA your own taste. It’s a simple little example, but everything from codebase structure to dashboard layout to how features in chatgpt fit together all fit under here. Their implementations are “correct”, the tests pass, I just hate the choices they made.

“The tests pass, I just hate the choices” sums up my experience with coding agents perfectly. We’re working on our first two apps for MacStories (more in the near future), and while I’m not writing the code, I have to spend so much time thinking about usability decisions, onboarding flows, iOS best practices, and “it just doesn’t feel right” issues that models haven’t been able to abstract yet. I have to wonder if, in a rush to maximize infinite token usage and build everything “because why not?”, perhaps the big labs may have ignored these other aspects of good software design. I strongly believe that there’s more to a great app experience than elegant code that passes tests.

Knowing how these things can change so quickly in this scene, I’m sure this post will be obsolete in a few months. Maybe even weeks. But at this moment in time, Davis’ article is spot-on and human decisions still make a dent in software coded by agents.

Permalink