This Week's Sponsor:

Inoreader

Boost Productivity and Gain Insights with AI-Powered Intelligence Tools

Posts in notes

Notes on Early Mac Studio AI Benchmarks with Qwen3-235B-A22B and Qwen2.5-VL-72B

I received a top-of-the-line Mac Studio (M3 Ultra, 512 GB of RAM, 8 TB of storage) on loan from Apple last week, and I thought I’d use this opportunity to revive something I’ve been mulling over for some time: more short-form blogging on MacStories in the form of brief “notes” with a dedicated Notes category on the site. Expect more of these “low-pressure”, quick posts in the future.

I’ve been sent this Mac Studio as part of my ongoing experiments with assistive AI and automation, and one of the things I plan to do over the coming weeks and months is playing around with local LLMs that tap into the power of Apple Silicon and the incredible performance headroom afforded by the M3 Ultra and this computer’s specs. I have a lot to learn when it comes to local AI (my shortcuts and experiments so far have focused on cloud models and the Shortcuts app combined with the LLM CLI), but since I had to start somewhere, I downloaded LM Studio and Ollama, installed the llm-ollama plugin, and began experimenting with open-weights models (served from Hugging Face as well as the Ollama library) both in the GGUF format and Apple’s own MLX framework.

LM Studio.

LM Studio.

I posted some of these early tests on Bluesky. I ran the massive Qwen3-235B-A22B model (a Mixture-of-Experts model with 235 billion parameters, 22 billion of which activated at once) with both GGUF and MLX using the beta version of the LM Studio app, and these were the results:

  • GGUF: 16 tokens/second, ~133 GB of RAM used
  • MLX: 24 tok/sec, ~124 GB RAM

As you can see from these first benchmarks (both based on the 4-bit quant of Qwen3-235B-A22B), the Apple Silicon-optimized version of the model resulted in better performance both for token generation and memory usage. Regardless of the version, the Mac Studio absolutely didn’t care and I could barely hear the fans going.

I also wanted to play around with the new generation of vision models (VLMs) to test modern OCR capabilities of these models. One of the tasks that has become kind of a personal AI eval for me lately is taking a long screenshot of a shortcut from the Shortcuts app (using CleanShot’s scrolling captures) and feed it either as a full-res PNG or PDF to an LLM. As I shared before, due to image compression, the vast majority of cloud LLMs either fail to accept the image as input or compresses the image so much that graphical artifacts lead to severe hallucinations in the text analysis of the image. Only o4-mini-high – thanks to its more agentic capabilities and tool-calling – was able to produce a decent output; even then, that was only possible because o4-mini-high decided to slice the image in multiple parts and iterate through each one with discrete pytesseract calls. The task took almost seven minutes to run in ChatGPT.

This morning, I installed the 72-billion parameter version of Qwen2.5-VL, gave it a full-resolution screenshot of a 40-action shortcut, and let it run with Ollama and llm-ollama. After 3.5 minutes and around 100 GB RAM usage, I got a really good, Markdown-formatted analysis of my shortcut back from the model.

To make the experience nicer, I even built a small local-scanning utility that lets me pick an image from Shortcuts and runs it through Qwen2.5-VL (72B) using the ‘Run Shell Script’ action on macOS. It worked beautifully on my first try. Amusingly, the smaller version of Qwen2.5-VL (32B) thought my photo of ergonomic mice was a “collection of seashells”. Fair enough: there’s a reason bigger models are heavier and costlier to run.

Given my struggles with OCR and document analysis with cloud-hosted models, I’m very excited about the potential of local VLMs that bypass memory constraints thanks to the M3 Ultra and provide accurate results in just a few minutes without having to upload private images or PDFs anywhere. I’ve been writing a lot about this idea of “hybrid automation” that combines traditional Mac scripting tools, Shortcuts, and LLMs to unlock workflows that just weren’t possible before; I feel like the power of this Mac Studio is going to be an amazing accelerator for that.

Next up on my list: understanding how to run MLX models with mlx-lm, investigating long-context models with dual-chunk attention support (looking at you, Qwen 2.5), and experimenting with Gemma 3. Fun times ahead!

Faking ‘Clamshell Mode’ with External Displays in iPadOS 17

A simple setting can be used as a workaround for clamshell mode in iPadOS 17.

A simple setting can be used as a workaround for clamshell mode in iPadOS 17.

Fernando Silva of 9to5Mac came up with a clever workaround to have ‘clamshell mode’ in iPadOS 17 when an iPad is connected to an external display. The catch: it doesn’t really turn off the iPad’s built-in display.

Now before readers start spamming the comments, this is not true clamshell mode. True clamshell mode kills the screen of the host computer and moves everything from that display to the external monitor. This will not do that. But this workaround will allow you to close your iPad Pro, connect a Bluetooth keyboard and mouse, and still be able to use Stage Manager on an external display.

Essentially, the method involves disabling the ‘Lock / Unlock’ toggle in Settings ⇾ Display & Brightness that controls whether the iPad’s screen should lock when a cover is closed on top of it. This has been the iPad’s default behavior since the iPad 2 and the debut of the Smart Cover, and it still applies to the latest iPad Pro and Magic Keyboard: when the cover is closed, the iPad gets automatically locked. However, this setting can be disabled, and if you do, then sure: you could close an iPad Pro and continue using iPadOS on the external display without seeing the iPad’s built-in display. Except the iPad’s display is always on behind the scenes, which is not ideal.1

Still: if we’re supposed to accept this workaround as the only way to fake ‘clamshell mode’ in iPadOS 17, I would suggest some additions to improve the experience.

Read more

Last Week, on Club MacStories: Symlinks for Windows and macOS, File Organization Tips, Batch-Converting Saved Timers, and an Upcoming ‘Peek Performance’ Town Hall

Because Club MacStories now encompasses more than just newsletters, we’ve created a guide to the past week’s happenings along with a look at what’s coming up next:

Monthly Log: February 2022

Metroid running at 4K on Federico's MacBook Pro

Metroid running at 4K on Federico’s MacBook Pro

MacStories Weekly: Issue 310

Up Next

Next week on Club MacStories:

  • On March 8th, at 5:00 pm Eastern US time, we’ll be holding a live audio Town Hall in the Club MacStories+ and Club Premier Discord community. Join Federico, John, and Alex for reactions to the day’s events and to ask any questions you may have. More details about the Town Hall are available in the Announcements channel on Discord.
  • In MacStories Weekly 311, John will publish a shortcut for tweeting links to web articles via Typefully.

Apple Music, Exclusive Extras, and Merch

Apple and Billie Eilish, whose highly anticipated album WHEN WE ALL FALL ASLEEP, WHERE DO WE GO? (out March 29) has set a new record for pre-adds on Apple Music, have launched an interesting new kind of partnership on the company’s streaming service. At this link (which is not the same as the standard artist page for Billie Eilish on Apple Music), you’ll find a custom page featuring an exclusive music video for you should see me in a crown, the upcoming album that you can pre-add to your library, an Essentials playlist for Billie Eilish’s previous hits, two Beats 1 interviews, and, for the first time on Apple Music (that I can recall), a link to buy a limited edition merch collection.

The merch drop is available at this page, which is a Shopify store with Apple Music branding that offers a t-shirt and hoodie designed by streetwear artist Don C, featuring Takashi Murakami’s artwork from the aforementioned music video. The purchase flow features Apple Pay support; both the website and email receipts contain links to watch the video, pre-add the album, and listen to the Essentials playlist on Apple Music.

For a while now, I’ve been arguing that Apple Music should offer the ability to buy exclusive merch and concert tickets to support your favorite artists without leaving the app. The move would fit nicely with Apple’s growing focus on services (you have to assume the company would take a cut from every transaction), it would increase the lock-in aspect of Apple Music (because you can only get those exclusive extras on Apple’s service), and it would provide artists with an integrated, more effective solution to connect with fans directly than yet another attempt at social networking.

This collaboration with Billie Eilish feels like a first step in that direction, with Apple actively promoting the limited edition sale and embedding different types of exclusive content (video, merch, Beats 1 interviews) in a single custom page. I wouldn’t be surprised if Apple continues to test this approach with a handful of other artists who have major releases coming up in 2019.

The Reliable Simplicity of AirPods

Chris Welch, writing for The Verge on AirPods’ advantage over other wireless earbuds:

AirPods are the best truly wireless earbuds available because they nail the essentials like ease of use, reliability, and battery life. There are alternatives that definitely_ sound_ better from Bose, B&O Play, and other. But they often cost more and all of them experience occasional audio dropouts. AirPods don’t. I’d argue they’re maybe the best first-gen product Apple has ever made. Unfortunately, I’m one of the sad souls whose ears just aren’t a match for the AirPods — and I’m a nerd who likes having both an iPhone and Android phone around — so I’ve been searching for the best non-Apple option.

But some 14 months after AirPods shipped, there’s still no clear cut competitor that’s truly better at the important stuff. They all lack the magic sauce that is Apple’s W1 chip, which improves pairing, range, and battery life for the AirPods. At this point I think it’s fair to say that Bluetooth alone isn’t enough to make these gadgets work smoothly. Hopefully the connection will be more sturdy once more earbuds with Bluetooth 5 hit the market. And Qualcomm is also putting in work to help improve reliability.

I haven’t tested all the wireless earbuds Welch has, but I have some anecdotal experience here.

A few months ago, I bought the B&O E8 earbuds on Amazon. After getting a 4K HDR TV for Black Friday (the 55-inch LG B7), I realized that I wanted to be able to watch a movie or play videogames while lying in bed without having to put bulky over-ear Bluetooth headphones on. Essentially, I wanted AirPods for my TV, but I didn’t want to use the AirPods that were already paired with my iPhone and iPad. I wanted something that I could take out of the case, put on, and be done with. So instead of getting a second pair of AirPods, I decided to try the E8.

I like the way the E8 sound and I’m a fan of the Comply foam tips. The case is elegant (though not as intuitive as the AirPods’ case) and the gestures can be confusing. My problem is that, despite sitting 3 meters away from the TV, one of the earbuds constantly drops out. I sometimes have to sit perfectly still to ensure the audio doesn’t cut out – quite often, even turning my head causes the audio to drop out in one of the E8. I’m still going to use these because I like the freedom granted by a truly wireless experience and because I’ve found the ideal position that doesn’t cause audio issues, but I’m not a happy customer. Also, it’s too late to return them now.

A couple of days ago, I was doing chores around the house. I usually listen to podcasts with my AirPods on if it’s early and my girlfriend is still sleeping, which means I leave my iPhone in the kitchen and move around wearing AirPods. At one point, I needed to check out something outside (we have a very spacious terrace – large enough for the dogs to run around) and I just walked out while listening to a podcast.

A couple of minutes later, the audio started cutting out. My first thought was that something in Overcast was broken. It took me a solid minute to realize that I had walked too far away from the iPhone inside the house. I’m so used to the incredible reliability and simplicity of my AirPods, it didn’t even occur to me that I shouldn’t have left my iPhone 15 meters and two rooms away.

The Cases for (and Against) Apple Adopting USB-C on Future iPhones

Jason Snell, writing for Macworld on the possibility of Apple adopting USB-C on future iPhones:

But the Lightning paragraph–that’s the really puzzling one. At first parsing, it comes across as a flat-out statement that Apple is going to ditch Lightning for the USB-C connector currently found on the MacBook and MacBook Pro. But a second read highlights some of the details–power cord and other peripheral devices?–that make you wonder if this might be a misreading of a decision to replace the USB-A-based cords and power adapters that come in the iPhone box with USB-C models. (I’m also a bit baffled by how the Lightning connector is “original,” unless it means it’s like a Netflix Original.)

Still, the Wall Street Journal would appear to be a more visible and reputable source than an analyst or blog with some sources in Apple’s supply chain. It’s generally considered to be one of the places where Apple has itself tactically leaked information in the past. So let’s take a moment and consider this rumor seriously. What would drive Apple to kill the Lightning connector, and why would it keep it around?

I’ve been going back and forth on this since yesterday’s report on The Wall Street Journal. Like Jason, I see both positive aspects and downsides to replacing Lightning with USB-C on the iPhone, most of which I highlighted on Connected. Jason’s article perfectly encapsulates my thoughts and questions.

USB-C represents the dream of a single, small, reversible connector that works with every device, and it’s being adopted by the entire tech industry. USB-C isn’t as small as Lightning but it’s small enough. More importantly, it’d allow users to use one connector for everything; USB-A, while universal on desktop computers, never achieved ubiquity because it wasn’t suited for mobile devices. USB-C is.

Conversely, Lightning is under Apple’s control and Apple likes the idea of controlling their stack as much as possible (for many different reasons). A transition to USB-C would be costly for users in the short term, and it would be extremely perplexing the year after the iPhone 7 fully embraced Lightning.

Furthermore, unlike the transition from 30-pin to Lightning in 2012, Apple now has a richer, more lucrative ecosystem of accessories and devices based on Lightning, from AirPods and Apple Pencil to keyboards, mice, EarPods, game controllers, Siri remotes, and more. Moving away from Lightning means transitioning several product lines to a standard that Apple doesn’t own. It means additional inconsistency across the board.

Like I said, I’m not sure where I stand on this yet. These are discussions that Apple likely has already explored and settled internally. I’m leaning towards USB-C everywhere, but I’m afraid of transition costs and setting a precedent for future standards adopted by other companies (what if mini-USB-C comes out in two years?).

In the meantime, I know this: I’m upgrading to USB-C cables and accessories as much as I can (I just bought this charger and cable; the Nintendo Switch was a good excuse to start early) and I would love to have a USB-C port on the next iPad Pro. If there’s one place where Apple could start adopting peripherals typically used with PCs, that’d be the iPad.

The TV App as a Supporting Actor

Joe Steel makes a good point in his look at this week’s Apple TV announcements:

Why is TV the app an app and not the Home screen on the device? It’s obviously modeled after the same ideas that go into other streaming devices that expose content rather than app icons, so why is this a siloed launcher I have to navigate into and out of? Why is this bolted on to the bizarre springboard-like interface of tvOS when it reproduces so much of it?

You could argue that people want to have access to apps that are not for movies or TV shows, but I would suggest that that probably occurs less often and would be satisfied by a button in the TV app that showed you the inane grid of application tiles if you wanted to get at something else.

As I argued yesterday on Connected, I think the new TV app should be the main interface of tvOS – the first thing you see when you turn on the Apple TV. Not a grid of app icons (a vestige of the iPhone), but a collection of content you can watch next.

It’s safe to assume that the majority of Apple TV owners turn on the device to watch something. But instead of being presented with a launch interface that highlights video content, tvOS focuses on icons. As someone who loves the simplicity of his Chromecast, and after having seen what Amazon is doing with the Fire TV’s Home screen, the tvOS Home screen looks genuinely dated and not built for a modern TV experience.

I think Apple has almost figured this out – the TV app looks like the kind of simplification and content-first approach tvOS needs. But by keeping it a separate app, and by restricting it to US-only at launch, Apple is continuing to enforce the iPhone’s Home screen model on every device they make (except the Mac).

That’s something the iPad, the Watch1, and the Apple TV all have in common – Home screen UIs lazily adapted from the iPhone. I wish Apple spent more time optimizing the Home screens of their devices for their different experiences.

  1. The Watch is doing slightly better than the other ones thanks to watchOS 3 and its Dock, but the odd honeycomb Home screen is still around, and it doesn’t make much sense on the device’s tiny screen. ↩︎

Spotify’s Release Radar is Discover Weekly for New Music

Release Radar's first take.

Release Radar’s first take.

Earlier today, Spotify unveiled Release Radar, an algorithmically-generated playlist updated Friday and designed to recommend new music. Like Discover Weekly, Release Radar tailors suggestions dynamically for your tastes, with the difference that it highlights newly released music from the past few weeks instead of anything you might be interested in. Essentially, Release Radar aims to be Discover Weekly for new song release.

The Verge has more details on how Spotify approached Release Radar after the success of Discover Weekly:

“When a new album drops, we don’t really have much information about it yet, so we don’t have any streaming data or playlisting data, and those are pretty much the two major components that make Discover Weekly work so well,” says Edward Newett, the engineering manager at Spotify in charge of Release Radar. “So some of the innovation happening now for the product is around audio research. We have an audio research team in New York that’s been experimenting with a lot of the newer deep learning techniques where we’re not looking at playlisting and collaborative filtering of users, but instead we’re looking at the actual audio itself.”

As a Discover Weekly fan, I think this is a fantastic idea. Discover Weekly has brought back the joy of discovering new music into my life, but the songs it recommends aren’t necessarily fresh. I can see Release Radar complement Discover Weekly as the week winds down with songs that I don’t know and are also new.

Already in today’s first version of Release Radar, I’ve found some excellent suggestions for songs released in the past two weeks. Spotify has their personalized discovery features down to a science at this point.

Conversely, I’m curious to see what Apple plans to do with their Discovery Mix feature of Apple Music announced at WWDC (shown here with a screenshot). Discovery Mix still hasn’t become available after four betas of iOS 10. I’m intrigued, but also a little skeptical.

Apple’s Data Collection in iOS 10

Ina Fried, writing for Recode, got more details from Apple on how the company will be collecting new data from iOS 10 devices using differential privacy.

First, it sounds like differential privacy will be applied to specific domains of data collection new in iOS 10:

As for what data is being collected, Apple says that differential privacy will initially be limited to four specific use cases: New words that users add to their local dictionaries, emojis typed by the user (so that Apple can suggest emoji replacements), deep links used inside apps (provided they are marked for public indexing) and lookup hints within notes.

As I tweeted earlier this week, crowdsourced deep link indexing was supposed to launch last year with iOS 9; Apple’s documentation mysteriously changed before the September release, and it’s clear now that the company decided to rewrite the feature with differential privacy behind the scenes. (I had a story about public indexing of deep links here.)

I’m also curious to know what Apple means by “emoji typed by the user”: in the current beta of iOS 10, emoji are automatically suggested if the system finds a match, either in the QuickType bar or with the full-text replacement in Messages. There’s no way to manually train emoji by “typing them”. I’d be curious to know how Apple will be tackling this – perhaps they’ll look at which emoji are not suggested and need to be inserted manually from the user?

I wonder if the decision to make more data collection opt-in will make it less effective. If the whole idea of differential privacy is to glean insight without being able to trace data back to individuals, does it really have to be off by default? If differential privacy works as advertised, part of me thinks Apple should enable it without asking first for the benefit of their services; on the other hand, I’m not surprised Apple doesn’t want to do it even if differential privacy makes it technically impossible to link any piece of data to an individual iOS user. To Apple’s eyes, that would be morally wrong. This very contrast is what makes Apple’s approach to services and data collection trickier (and, depending on your stance, more honest) than other companies’.

Also from the Recode article, this bit about object and scene recognition in the new Photos app:

Apple says it is not using iOS users’ cloud-stored photos to power the image recognition features in iOS 10, instead relying on other data sets to train its algorithms. (Apple hasn’t said what data it is using for that, other than to make clear it is not using its users photos.)

I’ve been thinking about this since the keynote: if Apple isn’t looking at user photos, where do the original concepts of “mountains” and “beach” come from? How do they develop an understanding of new objects that are created in human history (say, a new model of a car, a new videogame console, a new kind of train)?

Apple said at the keynote that “it’s easy to find photos on the Internet” (I’m paraphrasing). Occam’s razor suggests they struck deals with various image search databases or stock footage companies to train their algorithms for iOS 10.