Eshu Marneedi

The Compromises of Folding Phones

Eshu Marneedi — Fri, 31 Jul 2026 17:10:56 -0400

Dominic Preston, reporting for The Verge earlier in July:

A year after overhauling its Z Fold phone with a radically thinner design, Samsung has changed things up again, this time altering the shape entirely. The Z Fold 8 is wider than any of Samsung’s previous foldables, and shorter too, a size and shape somewhere between Samsung’s traditional Flip and Fold models.

In photos, I wasn’t impressed by the squat shape. In person, to my surprise, it feels great, and almost immediately intuitive. If you want to get a feel for it yourself, grab your passport — the Z Fold 8 is about the same height and less than half an inch more narrow.

This isn’t Samsung’s only new foldable this year, and in fact, its lineup just got more confusing. The Z Fold 8 may be the official follow-up to last year’s impressive Z Fold 7, but in truth the Z Fold 8 Ultra is the straightforward sequel. That phone has the same squarish inner screen as previous Z Fold phones and a similar triple rear camera. The name feels like a sly trick to get around this year’s inevitable price hikes: Samsung slapped “Ultra” on the box of its highest-end model to help cover for pricing the phone at $2,099.99, up $100 from the Z Fold 7.

I don’t have much interest in Samsung phones but nevertheless have been linking to Samsung Unpacked news for years now. What’s more intriguing to me is this new form factor, which is widely speculated to be the same as that of the folding iPhone due in about six weeks. The first generation of folding smartphones, beginning with the Galaxy Fold in 2019, had long, narrow front screens and square inner screens. This aspect ratio made both screens less than ideal to use: the front screen was too narrow, and the inner screen wasn’t wide enough to be used as a tablet. This new “squat” form factor sacrifices vertical length for width, making the inner screen much more useful while sacrificing day-to-day use of the front screen.

I think this is the way to go, and I remember saying so when the Pixel Fold was first announced in 2023. The entire point of a folding phone is the inner display, which should be large enough to eliminate the need for a tablet. But this great aspect ratio debate of the 2020s has proven something more important: folding phones are inherently full of compromises. The front screen — which most people will use most of the time — must be sacrificed to make the inner screen useful. This is still better than the old design, where both screens weren’t ideal to use, but it’s still a compromise. For the best experience, people will open up the phone like a book. This style of folding phones ought to be called “tablets with front screens as an afterthought.”

This brings me to the foldable iPhone. I’m positive the software will be world-class. App developers will be there Day 1 to ensure their apps work well on both displays; Apple will make the front display useful and practical. But this device is rumored to cost upward of $2,000, while forgoing the professional-grade cameras of the iPhones Pro, the pocketable yet useful display size of the standard iPhones, and the battery life of the iPhones Pro Max. (Just compare the specifications of the Galaxy Z Fold 8 with the Galaxy S26 Ultra.) I’m almost positive I won’t want to carry two displays around in my pocket, especially because I’ll have to open the phone up when I truly want a good experience. I’m almost positive I don’t want to spend $2,000 for worse cameras and battery life. And I’m almost positive I’m in the overwhelming majority.

I’ll refrain from passing judgment on the foldable iPhone until it’s announced, but I can judge the Galaxy Z Fold 8 — and I think it exposes too many compromises with this device class. None of this is Samsung’s fault, of course, but it does a great job illustrating why foldable phones will probably never be ready for primetime. There was a time many years ago when I was bullish on the folding phone, hoping technology would progress far enough that they’d be equivalent in price and features to their non-folding counterparts. That time has not come yet, and I’m not confident it ever will. Samsung and Apple have tried addressing the biggest drawback to folding phones, the form factor, this year. But the Galaxy Z Fold 8 proves that it’s just not enough.

‘You’ll Own Nothing And You’ll Be Happy’

Eshu Marneedi — Tue, 28 Jul 2026 16:50:01 -0400

Chance Miller, reporting for 9to5Mac:

Apple has officially announced Apple Upgrade, a new product leasing program available starting today in the United States.

Apple Upgrade, which is backed by Klarna, lets users lease iPhone, Apple Watch, Mac, and iPad purchases, with the option to easily upgrade at the end of the lease term. Here’s how it works.

Apple Upgrade is available across Apple’s product lines, including the latest iPhone 17 lineup, MacBook Air, Apple Watch Series 11, and much more.

iPhone and Apple Watch are available with 12-month and 24-month leasing terms, while Mac and iPad are available with 24-month and 36-month terms.

iPhone leases start at $17.99 per month for iPhone 17e

Apple Watch leases start at $11.99 per month for Apple Watch Series 11

Mac leases start at $24.99 per month for MacBook Air

iPad leases start at $15.99 per month for iPad mini

On the surface, this looks like a great deal. People can now lease any Apple product — including Macs with custom specifications — and upgrade their device at the end of the lease. They can also buy it, but then they’ll have to pay the remaining cost: Leasing a $2,000 MacBook Pro for 36 months only costs $39 a month, meaning someone would have to pay nearly $600 at the end of their term to own the product. At that point, you might as well just upgrade at no cost. These leases accrue no interest, either, so it seems like a win to lease a product like this and invest the money somewhere. That’s the pro-Apple Upgrade argument, at least.

I don’t like these leasing programs. I never used the now-discontinued iPhone Upgrade Program, despite buying a new iPhone every year, and this is the same reason why I encourage everyone I know to never purchase an iPhone through a carrier contract, even if they’re “handing them out for free.” There’s a monetary aspect to this — people living above their means; this program will make some people likely to purchase a device they can’t afford and become delinquent on their payments — but also a societal one. Why can’t we own anything anymore? Why is it that houses are so expensive that people are better off renting in most markets? Why are video games going digital-only, with software licenses that stipulate that you don’t actually own the games?

There are undoubtedly people who will benefit from this program. If Apple Upgrade lets you buy a computer that kickstarts your career, by all means, enjoy it. (Just please don’t buy a computer you couldn’t afford if push comes to shove.) But I take issue with not being able to own my smartphone — I don’t want a tool that only costs a thousand dollars to be tied to a lender who could theoretically take it away at any moment. That’s a dangerous precedent; at what point will everything we know and love be reduced to installment payments? Someone somewhere must draw the line. We have leases on cars because they cost tens of thousands of dollars — few people have that much in cash. But have we made it so far south in society that people need to lease computers?

This program rubs me the wrong way. Maybe it doesn’t for you — and that’s completely fine. But if it doesn’t, maybe you really will own nothing and be happy.

Apple Launches the MapKit for Automotive SDK, First With Ford

Eshu Marneedi — Fri, 24 Jul 2026 04:14:44 -0400

Apple Newsroom:

Apple and Ford today announced that Apple Maps will be integrated directly into Ford’s upcoming Universal Electric Vehicle (UEV) Platform through Apple’s new MapKit for Automotive SDK. Coming to Ford’s UEV Platform in 2027, this integration will deliver a beautiful and easy-to-use navigation experience powered by Apple Maps directly to the vehicle’s displays. Road-level Maps information will also enable Ford’s Latitude AI team to build a seamless hands-free driving experience.

Leveraging Apple’s new MapKit for Automotive SDK, Ford’s Universal Electric Vehicle Platform will offer drivers turn-by-turn directions using natural language, real-time traffic and incident information, intuitive search featuring detailed place cards, and routing options to help users easily find the best route to their destination. Additionally, Ford UEV drivers will benefit from an intelligent EV routing functionality, including battery preconditioning. Ford will also use road-level information from Apple Maps as part of the development of its next-generation BlueCruise hands-free highway driving capability to deliver a more seamless on-ramp-to-off-ramp experience.

Apple services have historically been sold as complements to Apple hardware: Apple Maps, for a decade, had been limited to Apple devices, replacing Google Maps. That is no longer the case; Apple TV is a canonical example of Apple’s interest in broadening the Apple services market. Similarly, this new MapKit SDK — short for software development kit — is an intentional expansion of Apple’s services market following the abject failure of CarPlay Ultra, a version of CarPlay meant to consume the entire car software experience. Even CarPlay’s most fervent supporters, like BMW and Ford, haven’t adopted CarPlay Ultra because it exerts too much control over a driver’s experience. The CarPlay Ultra gambit was clearly an effort to make something out of the Apple Car project, also a failure.

Ultimately, I think the CarPlay Ultra fiasco is indicative of the two opposing yet symbiotic forces within Apple: the product people and the services people. The product people want Apple to own as much of the “stack” as possible: the hardware, the software that hardware runs, and the services that software sells. They wanted to build a car, make software for it, and sell services through that software: that was Project Titan, the car project, in a nutshell. That is fundamentally the same strategy that governs the Apple Watch, MacBook Neo, even the iPhone. The services people, like those in charge of Apple TV and its content acquisitions like Formula 1 and Major League Soccer, don’t believe in that because they’d rather make money through advertisements and subscriptions. They paid hundreds of millions of dollars in upfront costs for MLS and F1 to sell more subscriptions, even to people who have never owned an Apple device.

The services people won here, largely because the product-driven CarPlay Ultra / Project Titan strategy — to own as much of the software experience as possible — failed. No car manufacturer was interested in that. Perhaps consumers would’ve been, since CarPlay by itself is overwhelmingly popular by Apple’s numbers, but it didn’t matter anyway because the difference in experience for iPhone users and non-iPhone users was too stark. (Not to mention, it was too much work for very little tangible benefit on the manufacturers’ end.) So now, Apple Maps — just like Apple Music in Tesla, General Motors, and Rivian vehicles — is a service built to make money independently. It doesn’t matter whether someone owns an iPhone or not, because they’ll still be forced to use Apple Maps in their Ford vehicle. And most importantly, they’ll be forced to look at Apple’s advertisements coming to Apple Maps soon.

The ads part is a tragedy because it signals a definite end to Apple’s product-driven mindset. Before Apple started putting ads in Apple services, they were also Apple products — you buy a subscription to Apple Music. Now, Apple services are partially ad-driven enterprises, where your use of the service itself draws revenue. In creating the MapKit for Automotive SDK, Apple doesn’t care about people spending money at the Apple Store or on in-app purchases. Apple isn’t delivering a product to consumers as much as it’s guaranteeing engagement to ad buyers. The consumers are not the primary consumers; corporations who pay Apple are. This is against the very philosophy Apple in many ways was founded upon: making high-quality products for consumers willing to pay premium prices. Modern Apple would probably object to the characterization that ads in Apple services cheapen the experience, but it’s true: ads in Apple Maps do not make for a premium experience.

The new Apple doesn’t seem to care about that. It’s uninterested in delivering a premium experience, and it’s uninterested in asking for your money to deliver that experience. Ergo, the MapKit for Automotive SDK.

U.S. Threatens to Sanction China Over AI Distillation

Eshu Marneedi — Wed, 22 Jul 2026 00:22:34 -0400

Ashley Capoot, reporting for CNBC:

U.S. Treasury Secretary Scott Bessent on Tuesday said the Trump administration will look into whether Chinese artificial intelligence models have been distilled from American models, stating that the government does not support “IP theft.”

Chinese open-weight models are gaining steam against leading offerings from American companies like OpenAI and Anthropic, sparking concerns from tech executives and government officials about the durability of the U.S. lead in the AI race. Moonshot AI, a Chinese startup, released a model called Kimi K3 earlier this month that outperforms those companies across some industry benchmarks. Open weight refers to models whose final trained parameters are publicly released for download, while the underlying code and training data remain private.

“If we see, especially that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft,” Bessent told Fox Business’ “Mornings with Maria” on Tuesday.

Bessent said the technical term for this theft is called distillation, which is an AI training method where a smaller, less capable model is built using outputs from an existing, stronger model. Anthropic sent a letter to the U.S. Senate Committee on Banking, Housing, and Urban Affairs last month alleging that the Chinese tech company Alibaba had carried out the “the largest known distillation attack” against it to date.

If distillation — training on data that doesn’t belong to the Chinese labs — is a sanctionable offense, then why isn’t pre-training? How is it fully legal for Chinese companies to use training data published by American citizens on the internet? This whole argument is bogus, and it’s quite evident that Bessent knows very little, if anything, about what he’s talking about. AI models are built on the idea that training a computer on the open internet is fair use — that pre-training is transformative enough under fair use doctrine, not “stealing.” American companies can train their models on Chinese text and vice versa because pre-training is inherently transformative, an argument I’m quite sympathetic to. I think that if we built a time machine back to 2022, a permission or compensation structure should’ve been created, but alas, this is the system we’re stuck with. The world’s copyright system is not suited for AI pre-training.

Back to the Bessent argument. Anthropic, and perhaps OpenAI, are perturbed that Chinese models — like Moonshot’s Kimi K3 open-weight large language model — were trained via distillation. Kimi K3 is cheaper than Claude Opus 4.8 and GPT-5.6 Sol, yet almost matches GPT-5.6 Sol in overall intelligence and beats Opus 4.8 entirely. Kimi K3 comes just shy of Claude Fable 5, a model over three times the price. While Anthropic can’t even be compelled to offer Fable 5 access to some of its paying customers, Kimi K3 offers similar performance at a fraction of the cost — but it was built by distilling Anthropic’s models. Anthropic’s view, which it has surely expressed to the government, is that this is an adversarial nation subsidizing AI development, stealing the secrets of American AI labs, and building a compelling product for much cheaper. From that perspective, Moonshot and China look guilty.

But that’s a pacifist view; Anthropic is wailing for pity from a government that has treated it abysmally. Just Tuesday, a judge approved a billion-dollar settlement between Anthropic and a litany of authors after it pirated books to train its models on — is that not “stealing?” Is Anthropic not one of the most blatant copyright infringers of the modern era? It’s quite obvious here that Anthropic is crying wolf not because China gained some unfair advantage in its training process, but because Anthropic is so upset that it wasted perhaps billions of dollars in subsidized tokens all for its users to prefer a foreign competitor’s clearly superior model. The problem here was never about morality or copyright or distillation; the AI labs are morally bankrupt, and they know that. If they didn’t, they wouldn’t be cutting billion-dollar checks to authors.

So where do we go from here? I think the solution is quite obvious: Anthropic must start distilling its own models. Train a massively expensive competitor to Fable 5, distill a new version of Opus on that model’s reasoning traces, and ship the new Opus at a much lower cost. We’re reaching a point in the AI industry where cost is a massive factor in the models enterprises and individuals use, as reasoning traces become more expensive. Yes, pre-training is currently the most expensive part of releasing a new model, but we’re already seeing the shift toward inference and post-training. New model releases nowadays are not new pre-training runs; GPT-5, the last “major” AI model from OpenAI, was a heavily post-trained GPT-4o from spring 2024. As LLMs mature and reasoning tokens become more valuable, inference will be the biggest operating cost, not pre-training. Train one model and distill three others, then offer inference at a competitive price. Distillation — from China and by the American labs themselves — is in the AI industry’s best interests. It saves on pre-training costs and leaves more room to make healthy margins on inference.

Apple Is Looking Into Buying AI Chip Companies for Private Cloud Compute

Eshu Marneedi — Sun, 19 Jul 2026 01:50:02 -0400

Marcus Mendes, reporting for 9to5Mac:

According to The Information, Apple has been exploring potential acquisitions of semiconductor startups to “boost its efforts to build server chips for running AI.”

From the report:

In recent months, the iPhone maker has talked with bankers about possible deals. It has also approached semiconductor startups to gauge their interest in selling themselves, the people said. Apple’s hunt for chip acquisitions comes as the company struggles with the performance of its own internal AI servers, which currently run on internally designed M2 Ultra chips.

The Information says Apple’s next-generation server chip, code-named Baltra, has slipped past its planned 2026 debut. In the meantime, Apple is using M2 Ultra-based systems for some of its own AI processing, while the more demanding tasks are handled by the Gemini-based model powering the new Siri, which runs on Nvidia GPUs in Google Cloud.

A big acquisition would be a departure from Apple’s usual M.O. of buying smaller startups from time to time, but it wouldn’t be unprecedented.

It would indeed be unusual for Apple to purchase a Silicon Valley start-up specializing in artificial intelligence processors, when Apple arguably makes the best consumer AI chips. As Bloomberg has reported extensively, Apple’s consumer processor strategy — for high-end Mac models, at least — has shifted considerably toward AI. People like buying Mac Studios and Mac minis to do local AI inference — they’ve become popular as cloud computers and mini data centers. Per The Information’s reporting, Apple’s internal processors are also on that trajectory, albeit delayed for unclear reasons. Apple’s processors aren’t just fast and capable of AI inference, but they’re extremely efficient, which makes them relatively inexpensive to use in the long run. It’s no wonder why Apple has retaken the No. 1 market capitalization, surpassing Nvidia on Friday.

Regardless, this news is quite interesting because it proves Apple is intent on shipping more powerful, capable models in the future — perhaps ones that can write code or perform agentic tasks. It wants people to be able to run those models on their own devices — hence why the company is testing Mac Studios with 1.5 terabytes of memory and skipping high-end M6-generation chips entirely — and to eliminate its reliance on Google Cloud, or maybe even Google altogether. If Apple were to buy such a company, it could probably pre-train its own models: a lack of compute was one of the key limitations that led to the Google partnership. Apple’s new leadership doesn’t see AI as a fleeting trend. It is finally taking it seriously, something we couldn’t say about the AI team under John Giannandrea, the machine learning chief who was stripped of his responsibilities and replaced by Mike Rockwell, the executive previously in charge of Apple Vision Pro.

I think Apple should lean into efficient, low-cost AI inference on both the hardware and software fronts. Siri AI clearly emphasizes on-device processing, and the company’s future Apple silicon plans do, too. Readers of this blog know that my single biggest qualm with the AI industry for now is that its economics aren’t rooted in logic. Anthropic is compute-constrained, so much so that they’re under heavy competition from Chinese AI labs that can offer inexpensive tokens and relatively the same performance as Anthropic’s best models. OpenAI is deathly unprofitable, leading some pundits to believe that its economics will spell its demise. Compute is the biggest story of Silicon Valley this summer, whether it’s skyrocketing memory prices for consumer products or unusual deals like Meta selling compute to Anthropic.

Whatever happens here — if Apple chooses to buy a compute start-up or not — the fact that it is even entertaining the idea is proof that it sees Apple silicon as its golden ticket in the AI space. As open-weight models get smaller and more powerful, people will buy more powerful Macs marketed as the best for AI inference; as the frontier AI labs become more compute constrained — and their flawed economics catch up to them — people will look to Apple’s models and products. Siri AI is merely the beginning of the Apple AI story.

A Bluetooth Smart Speaker Will Not Fix OpenAI’s Problems

Eshu Marneedi — Thu, 16 Jul 2026 01:16:54 -0400

Mark Gurman, reporting for Bloomberg:

OpenAI’s much-anticipated push into consumer devices is slated to begin with a mobile, screen-free smart speaker designed to be a new type of home computer for the AI era, according to people familiar with the matter.

The product — still under development — is meant to serve as a humanlike AI companion that lives in the home, said the people, who asked not to be identified because the project hasn’t been announced. It will help control smart-home appliances, play media, answer questions, respond to messages and tap into the range of capabilities offered by OpenAI’s ChatGPT, they said.

Though the new product resembles a speaker, OpenAI internally describes it as the first of its kind: a computer built for AI to help make busy people more productive. It includes a camera and other sensors that help it understand a user’s surroundings and context, as well as advanced AI models beyond those available on conventional smart speakers.

Remember Jibo? Jibo, developed by a professor at the Massachusetts Institute of Technology after a successful Kickstarter campaign, was an animated voice robot that would move around with a motorized base and use its camera to detect facial expressions and gestures. Jibo came years before the modern large language model, and its use was limited, but I’d imagine this device from OpenAI to be much like it. Like Jibo, this is just not a breakthrough consumer product. It’s a Bluetooth smart speaker with a motorized base — which Gurman describes as “mechanical elements that can move on their own” — and ChatGPT built in.

You can already buy a smart speaker with an LLM for just $100: it’s the Google Home Speaker, and — news flash — it’s not a revolutionary product. It’s certainly not revolutionary enough to warrant buying Jony Ive’s hardware company and poaching trade secrets from Apple. But none of this even considers how unfeasible it will be to have a hardware product from OpenAI — a company almost universally taken with distrust — listening and peering into people’s homes all day long. I’d be hard-pressed to find a single person who wants that. There are people on the streets irate about OpenAI data centers being built in their backyards; on what planet will those people put an OpenAI product in their homes?

OpenAI has spent billions of dollars on this hardware venture. It has upended the technology industry — it has decimated entry-level jobs, transformed computer science education, and sowed doubt in tech as a whole. People are down on artificial intelligence and technology writ large. For this product — and OpenAI as a company — to be successful, it must transform not only its public image, but that of the technology it wants to sell. I have a hard time believing a Bluetooth smart speaker designed by Ive with ChatGPT built in is a serious solution to that problem. Maybe it’s the most beautiful speaker in the world. Maybe it’s made with gold or some opulent frills. That doesn’t change the fact that OpenAI has a gargantuan favorability problem.

OpenAI also has a money problem. People do not like paying for AI. Only a small fraction of ChatGPT users are paying subscribers, and most of those people are on the lowest subscription tiers. The people who do pay heavily for AI burn subsidized tokens so hard that Sam Altman, the company’s chief executive, is holding back on an initial public offering until the company’s finances are more stable. What company that is struggling to sell consumers on a $20 subscription and enterprises at base rates for its models could possibly think it could sell hardware? The one popular product OpenAI makes, ChatGPT, is not profitable. It can’t even persuade people to pay for its flagship innovation, let alone a designer smart speaker.

I really don’t want to be Steve Ballmer. “$500? Fully subsidized? With a plan?” But I have a feeling this take will fare much better than my take from last year. Over a year later, Apple doesn’t look bad anymore; OpenAI does.

Hands-on With iOS and iPadOS 27, macOS 27 Golden Gate, and Siri AI

Eshu Marneedi — Mon, 13 Jul 2026 16:51:06 -0400

A return to form and function

Image: Apple.

In 2009, at Apple’s Worldwide Developers Conference, Bertrand Serlet walked onstage and presented an audacious claim: “Zero New Features,” the slide read.

Serlet, Apple’s then-senior vice president of software engineering, was introducing OS X 10.6 Snow Leopard, a release Apple claimed would emphasize small “refinements” and major updates to internal technologies. But to the public, it really was “Zero New Features” — a quote forever cemented in Apple’s modern history as one of its most notable. As Serlet later explained, OS X Snow Leopard had plenty of new features, most notably Grand Central Dispatch, which brought multithreading to OS X for the first time. OS X Snow Leopard established the architecture on which all personal computers, including the iPhone and the Apple Watch, would eventually rely. Its stability solidified Apple’s reputation for reliable, high-quality software, so much so that the coming decade of pervasive, sloppily thought-out user interfaces — beginning with OS X 10.10 Yosemite — was treated as an aberration.

In early June, Apple’s first real announcement at WWDC was neither novel nor revolutionary. It was not a coding agent that would put all white-collar workers out of jobs by 2030, augmented reality glasses, or even as much as a tab bar in the Photos app — it was a slider to adjust the translucency of Liquid Glass. Then came the stunning admission that macOS 26 Tahoe’s window corner radii were too inconsistent and radical: Apple would decrease and equalize all corner radii throughout the system in macOS 27 Golden Gate. Liquid Glass app icons would use new glass layers for a more polished, legible appearance. The central processing unit scheduler — the program that tells the CPU when to execute a process — was made more efficient. The Spotlight index was rewritten, and the Mail app’s search function was markedly improved. These were among the first “features” on which Apple spent the precious opening minutes of one of its most-watched presentations, and it was marvelous.

All of Apple’s releases this year converge on a single philosophy: returning Apple to its original promise of reliable yet innovative software. And much as OS X Snow Leopard — the company’s last true “cleanup year,” with hundreds of bug fixes and underlying performance improvements — became known for transforming personal computing for decades to come, so will iOS 27 and macOS Golden Gate. After a section of the keynote aimed, candidly, at regulators keen to scrutinize Apple’s child safety efforts, Apple introduced what I’d loosely call the Grand Central Dispatch of WWDC 2026: Siri AI, a completely reworked, large language model-powered version of Siri promised two years ago. After years of rumors, personnel changes, and bottlenecks, the “more personalized Siri” is finally here, and it’s both a Herculean reworking of the internals and a breakthrough consumer product.

I have spent over a month with all the new operating systems and Siri AI, and I can confidently report that Apple Intelligence, or at least this new version, has changed the way I use Apple products. It is the first artificial intelligence system that feels personal, like it knows me and is willing to use the decades of information I have stored on my Mac and iPhone. It sheds the gimmickry of current AI systems and, dare I say, makes you feel like you’re living in the future. Apple has proven that the answer to modern computing’s biggest questions comes not via the LLM alone, but via the cowling built around the model. Apple is a product company, and Siri AI is the best LLM product, engineered by people who astutely know what makes a great one. It’s an ice-cold glass of water in the hell of coding agents that will supposedly assume everyone’s jobs by the decade’s end.

But the “Snow Leopard” nature of this year’s operating systems leaves little to write about. There is, of course, Siri AI, coupled with some interesting changes to Liquid Glass, but there are almost no new features in this year’s releases. That’s OK — more than tolerable, in my book. I can ostensibly name dozens of features Apple announced in just a few years that occupied collective hours of keynote time, but hardly anyone uses: StandBy, Image Playground, tinted app icons, the Journal app, and Stage Manager, as pertinent examples. It’s not even that these features are useless, per se, but that they’re just too onerous to find. This was, in fact, the lede of my OS review from two years ago: Apple adds dozens of new, marquee features each year to go undiscovered and unused, while the core parts of the OS languish. Perhaps hundreds of millions of people search for emails in the Mail app, and it has been terrible for years. It isn’t anymore.

This is to say that my relative brevity in this year’s review should not be considered a gripe, but great praise. The adage of “comparison is the thief of joy,” begone — this year’s releases are a perfect replica of what made OS X Snow Leopard one of the most beloved Apple platform releases in history. They’re reliable yet teetering-on-the-edge-of-revolutionary pieces of software, and they’ll undoubtedly change the way millions of people use their devices in the fall. And I think a throwback to 2009 — a mélange between innovation and reliability — is the ideal place to be in a time of grave uncertainty and mistrust in the technology industry.

Siri AI

Siri AI is Apple’s marketing name for this year’s preeminent feature: a generative artificial intelligence-powered version of Siri that can search the web, perform actions in apps through App Intents, and peer into the personal context — a collection of personal data from first- and third-party apps. At its core, Siri AI uses a suite of on-device and cloud models to accomplish these three primary objectives. For more detail on how these models work in tandem, read my initial reactions to the Siri AI announcement from June. An on-device model, called the system orchestrator, chooses which model is best suited for the query and what information it needs, such as the personal context, an App Intent, or a web search. The model then uses the required tools to do the requested work and returns an answer.

This admittedly oversimplified description may be reminiscent of a chatbot, like ChatGPT, Claude, or — especially — Gemini. But over my weeks with Siri AI, I think the product is far closer to an interactive assistant than to a chatbot. Products like ChatGPT have taught us over nearly four years that natural-language input is the future of computation. Older virtual assistants, like the non-AI Siri and Google Assistant, used a limited understanding of human language to provide deterministic, reliable answers. They could neither process language naturally nor generate language of their own; they were instead hand-trained to recognize certain phrases like “What’s the weather tomorrow?” They were handed a predefined set of tools and were essentially told to match certain queries to them. In Siri’s case, the assistant only generated preprogrammed outputs, handwritten by Apple employees for different scenarios.

All of this is to say that those assistants were only useful for a handful of tasks, usually ones performed on-device. Siri — to this day, in non-beta versions of the operating systems — types almost all knowledge-based queries into Bing. It has virtually no knowledge of its own. Chatbots lie on the opposite end of the virtual-assistant spectrum: They claim to hold the world’s knowledge and have such a grasp on it that they can write complex computer programs from scratch or solve problems that have stumped mathematicians for decades. They’re mathematical models of human language, designed to predict the answer to a question nondeterministically. Siri AI lands somewhere in the middle of the spectrum, still unquestionably best suited, like its prior versions, for personal commands, yet with the breadth of knowledge that inherently comes from an LLM foundation. Siri AI merely leverages an LLM to realize the original product’s vision from 2011 — it is unapologetically not a chatbot, despite using the same underlying technology.

I say this for two reasons: the models and the interface. The former is straightforward: The models are nothing like GPT-5.6 Sol from OpenAI or Claude Fable 5 from Anthropic. AI enthusiasts use terms like “big model smell” to refer to the breadth of knowledge these frontier models have, and it’s immediately apparent upon interacting with Siri AI that “big model smell” was never Apple’s intention. Siri AI speaks organically, in short sentences, and uses little formatting. It prefers to quote much of the source material, such as when searching the personal context or the web. The model seldom launches into long reasoning traces and usually generates laconic responses. It’s lightning-fast — almost latency-free, to my delight — in nearly every interaction, choosing to process most queries on-device. Those are not characteristics of frontier-level models, which write with a certain cadence that Siri AI cannot match. Siri AI would clearly struggle to answer first-year calculus problems, let alone discover novel mathematics or write code; it is neither amusing nor particularly useful as a conversational partner.

The latter reason is, broadly, the rest of this section. People can open Siri AI on iOS in four ways: pressing and holding the Side Button to launch voice mode, using the Siri app, swiping down from the center of the Home Screen to open Spotlight, or swiping down from the top center of any app to launch the same Spotlight interface. The point is that Siri AI is available everywhere, including in apps. On macOS, people can use Spotlight or the “Hey Siri” wake word. Spotlight and Siri are not the same; rather, Spotlight leads to Siri. Spotlight works normally without any AI features unless the query no longer matches any results, in which case it offers to send it to Siri. If you prefer to talk to Siri directly, the voice mode bypasses Spotlight entirely, and so does the new, bespoke Siri app, which aggregates Siri conversations across devices through iCloud. Searching for “Siri” in Spotlight launches the Siri app.

Siri AI across modalities.

All of Siri AI’s new modalities use a beautiful semitransparent gradient; the traditional voice mode uses a new glassy orb. When using Siri AI from anywhere but the Siri app, it initially displays a short answer no longer than a paragraph. Manually swiping down on the initial result opens a mini window with a longer answer and, for select queries, web images. You can open the conversation in the Siri app from this window using a button at the top trailing edge. This reveals how Apple envisions Siri AI: It’s for direct answers, not long chats. People don’t even need to interact with the Siri app at all if they don’t want to — the app functions more as a place to recall past conversations than to start new chats. It is, of course, still possible to start chats and voice conversations from the Siri app — because anything else would be unintuitive — but that’s clearly not the app’s primary purpose. The Siri app is not a competitor to ChatGPT and isn’t a place where people will spend significant time.

The Siri AI app in iOS 27.

In practice, I start almost all of my Siri conversations on iOS through voice mode¹ — when I’m not in public, at least — and on macOS through Spotlight, just because speaking to a desktop computer seems daft. I scarcely even open the Siri app because it’s quite barebones, at least for now. It is neither as visually appealing nor as useful as the ChatGPT app, and offers only basic file uploads and search. I especially love all of the new animations and sounds when Siri is working: a circle of dots rotates as Siri reasons through an answer, and results drop down gracefully from the Dynamic Island on iOS. Results don’t stream in — Siri displays them in full only after it finishes writing the response. Siri AI is such a gorgeous, whimsical interface — it’s the canonical example of how best to use Liquid Glass. I also seldom go back and forth with Siri, except for clarifying questions, unlike my use of ChatGPT or Claude. Siri doesn’t invite conversation and feels more like a one-and-done assistant than a robot worth talking to or reasoning with.

The interface of Siri AI.

My most frequent use case for Siri AI is simple, fact-based questions. The model does occasionally hallucinate², and I wouldn’t use it for any serious work — such as fact-checking an article or learning a new topic — but it’s great for off-the-cuff questions. Some recent questions I’ve asked Siri AI include:

“When did Delta Air Lines take delivery of its first Airbus A350?”
“How many people voted for Kamala Harris in 2024?”
“Has India ever qualified for the World Cup?”
“Who won the Grammy Award for Best Alternative Music Album in 2024?” And then, “What about 2023?”

Siri answered each of these questions accurately, providing sufficient and reliable references for every one. These are all questions the older Siri would either refuse to answer — kicking off a Bing search instead — or send to ChatGPT, which would respond with varying degrees of success. The ChatGPT integration was also one-way: I could not follow up within the Siri interface, and chats started within Siri would usually take a few hours to appear in the ChatGPT app. And sure, I could use the ChatGPT app for these questions, but Siri is instant, almost to an uncanny degree. It writes a sentence or two, occasionally shows an image, and links to the site at the bottom of the response for easy access. Siri AI is simply better than ChatGPT for trivial general-knowledge questions, thanks to the way it presents information.

This is not to say that Siri AI never makes mistakes. The fast on-device models Siri AI opts for most are woefully out of date and, frankly, often incompetent. It all comes down to the basic premise I outlined at the beginning of this section: Siri AI is not a chatbot. I have worked through many projects with Claude and ChatGPT, but Siri AI was unable to pick up on any of them. I just found it useless for even basic chatbot-style back-and-forth. For example, I was combating some awful Chinese bot traffic on this website early in June and wanted to set up a Cloudflare filter to screen visitors. Siri AI pulled old, unhelpful documentation from some shoddy AI-generated webpage; ChatGPT told me exactly how to do it and what to do if it failed. Siri AI is an LLM-powered digital assistant down to the little interface details: the mini window, its brevity, the web references, and the integration with Spotlight. It would be an assistant first and foremost by design even if it used frontier-class models.

The fortunate news is that Siri AI is the best digital assistant available. The crème de la crème of the new architecture — and what propels Siri AI from a mediocre ChatGPT alternative of three years ago to one of the most elegant, transformative pieces of software — is the personal context, derived from Apple’s semantic index. Apple did not bury the lede about where the semantic index comes from: Siri and Spotlight are joined at the hip because the Spotlight index and the semantic index are the same. When Siri AI is first enabled on a new device, Spotlight takes up to a week to index files, notes, emails, and much more. The personal context exposes this information to Siri AI, including data from third-party apps that opt to donate to the semantic index. If Siri determines that it must use a particular note in Apple Notes, the system orchestrator will find the note in the personal context and provide it to the model for processing.

It’s self-evident why Spotlight indexing takes so long after updating to iOS 27: The system must index virtually every bit of information available across every app with a high degree of specificity. If it needs a text message from a decade ago to answer a question, the system should be able to find it in reasonable time. Siri AI falls apart without this functionality. And the model should be able to query the personal context with natural language — the system would be useless if it required precise search terms, hence the use of a semantic index rather than the lexical Spotlight index. If a certain Bloomberg reporter ever compiles The Siri Files from the past two years, I’d imagine the personal context would be known as the most complex part of the architecture. It is legitimately flawless in every test I’ve tried — iOS just seems to know me better than I do. In practice, Siri was able to find messages that even the Messages app’s search function couldn’t, notes whose titles I didn’t know, and poorly formatted emails from years ago.

This is only possible due to Siri AI’s LLM architecture. As I wrote earlier, the benefit of an LLM is not purely that it has a lossy recollection of the world’s knowledge, but that it has a grasp of human language. Siri AI captures intent better than any modern AI system — it was tastefully pre-trained to understand nuance and sometimes incomprehensible speech. “Find that note about note-taking apps in 2019-ish, maybe August.” This is a legitimate transcript of a Siri conversation I had, and the system found the note in mere seconds. It was actually from August 2020, and the note didn’t mention the phrase “note-taking apps” once. Another request: “Remind me to water the plants before my appointment tomorrow.” I had never put the appointment on my calendar — Siri found the appointment in my email after noticing it wasn’t in Calendar and set the reminder 30 minutes beforehand. The system is both persistent and clever, wonted qualities of a great assistant.

A perennial frustration that all iPhone users have likely had with Siri at least once is that talking to it felt like walking on eggshells. The new Siri AI is remarkably adept at understanding the messiest questions, whether it’s searching for a bit of niche information online, recalling the personal context, or calling an App Intent. There’s no longer a need to ask in a particular format: Requests like “What’s 4 p.m. England time?” are interpreted well. Siri AI understood I was asking for the conversion from 4 p.m. British Summer Time to Eastern time; the old Siri would downright give up (“I’m not sure I understand”) or just state the current time in England for such a query. Telling Siri, “I watered the plants,” instructs it to mark the “water the plants” reminder as completed. “I watered the shrubs” also works, even if the reminder only has “plants” in the title. If Siri AI ever needs clarification — such as if dictation couldn’t capture the full sentence — it proactively asks: “Just to confirm, did you mean X or Y?”

Siri AI has some interesting limitations on the Mac; namely, it doesn’t have full access to the file system and shell, unlike AI coding agents. I think this is a baffling limitation, especially because Siri AI could become the de facto tool for people — particularly power users likely to use the Mac for development work — to tweak settings on their computers. I use Claude Code all the time to modify property list files, organize and rename files on my desktop, or find things to delete. Just the other day, I asked Claude to find which “Developer Tools” were occupying 100 gigabytes on my disk — Claude found the deprecated simulators I had installed years ago and deleted them for me. I realize this is a niche request, but I think it’s important for an agent like Siri AI to interact with the file system, similar to Claude Cowork. I could see many people using such functionality to clean up their computers and organize files, especially if it were in the safe, trusted hands of Apple software. Siri AI is limited to finding files and cannot move or delete them. It also can’t create or edit files, nor can it use AppleScript, a legacy scripting language used to automate some apps. It’s a frustrating impediment.

Siri AI on the Mac.

Siri AI instead has an astute on-screen awareness capability. Out of curiosity, I opened a photo from the television show “The Office” — the one where Pam Beesly says, “They’re the same picture” — and asked Siri, “What episode is this from?” Siri responded in seconds: “The image features the character Pam Beesly (played by Jenna Fischer) in a scene from the Season 7 finale, ‘Search Committee.’” It even cited four separate internet sources, including a template of the infamous meme. The speed at which the model retrieved the data is due to the innate efficiency of an on-device model, but having it embedded within the system in this fashion is exceedingly helpful. On-screen awareness can analyze the full contents of what’s open — like a full article or note that trails off-screen — and goes beyond sending a simple screenshot to Siri. Siri AI also knows about currently playing media, even from third-party apps that don’t yet have App Intents.

Siri AI’s on-screen awareness feature.

When Siri must perform actions within apps — such as marking a reminder as complete, as opposed to listing unfinished reminders — it uses an App Intent. An App Intent is a small piece of code that, when run, can quietly launch an app in the background to perform a task. The intent works as a bridge between the app action (e.g., marking a reminder complete) and the system, which calls that action. App Intents are ubiquitous in iOS and macOS today: They’re required for Shortcuts and interactive widgets, like those in Control Center. Siri AI can now call these intents with better precision and semantic understanding — it can write complete notes with paragraphs and formatting or create usefully named reminders. When Siri uses an intent, a small “window” into the app’s content appears — such as a list of reminders or a new note — indicating a particular app was used. Broadly, Siri can now fully take action within apps on a user’s behalf, realizing — to a certain promising extent — Silicon Valley’s end goal of agentic experiences. The only catch is that every action in every app must be an intent to be usable by Siri.

Most native apps have rich support for App Intents in the beta. For example, I asked Siri to “convert this dashed list to bullets” with a note open in Apple Notes, and it obliged. Another time, Siri summarized an article I was reading in Safari and included a link and the author’s name in a note. Requests like these required a mix of on-screen awareness, personal context retrieval, and an App Intent to write to that note. I’ve generally found Siri’s ability to find relevant documents and match my formatting and writing style to be one of its best use cases, though the model occasionally outputs unformatted Markdown or misaligned paragraphs³. Siri also chooses, as best it can, which app’s intents to call without manually specifying, but currently defaults to using first-party apps. I must specify every time that Siri should add a reminder to Things rather than Reminders, for example. I hope these preferences eventually sync through the personal context.

Triggering App Intents and the personal context from Siri AI.

Developers contribute to the personal context by using App Schemas to mark certain information as indexable and add app actions through Intent Schemas; they can tell the system about open views and documents through NSUserActivity and App Entities. I think adopting these technologies will be table stakes as iOS 27 rolls out to the public. In 2024, I feared a lack of third-party App Intent support would hamstring the Apple Intelligence initiative after Apple Vision Pro suffered the same fate. But after using Siri AI, I really believe most users will eagerly await support for it in all of their favorite apps, including corporate developers like Google and Meta. Independent apps like my favorites will probably have support on Day 1. Uber, for instance, does indeed have an incentive to get people into the Uber app, but it has a strong record of using native iOS features, like Live Activities. The Uber app today has an integration with Siri to quickly book a ride, and ride-share information from the Uber app is listed in Apple Maps. These are voluntary integrations Uber has decided are good for its business, and rightly so. I think if Uber doesn’t support Siri AI within a few months of its launch, it’ll be at a disadvantage — I expect nothing less from the company that shipped a ChatGPT app almost immediately.

Convincing corporate developers to donate data to the personal context will be a tougher sell, though more plausible than it was in 2024. Meta’s apps, like Threads, WhatsApp, and Instagram, have already adopted Liquid Glass, though they undoubtedly use custom interfaces that are a hassle to rip out and redesign. People weren’t even particularly clamoring for Liquid Glass redesigns of these apps — the sentiment around the redesign is generally neutral to negative — but Meta spent time doing it. Apple still has a chokehold on app design and UI paradigms, and it would be Meta’s loss not to give the personal context access to Instagram direct messages or WhatsApp chats. If anything, wouldn’t that encourage more people to use third-party messaging apps, since they’d act indistinguishably from Messages? Companies like Meta and Spotify have lobbied governments for decades for the same access to system experiences as first-party apps — why wouldn’t they take advantage of it now that it exists?

Perhaps this is naïve optimism, but over my month with Siri AI, it legitimately felt indispensable. Through clever product direction, Apple has leveraged the candidly mediocre LLMs from the Google partnership to build one of the most elegant, user-friendly, and useful AI products from Silicon Valley in the last four years. I find myself gravitating toward Siri AI for all sorts of small tasks I was typing into Google or using ChatGPT for: quick facts, image lookup, and basic web searches. Local AI is Apple’s power play, and this makes Siri the most useful, delightful, and convenient AI assistant soon to be available on the market.

But the genuinely innovative personal context and App Intents might just be the panacea Apple needs not only to catch up in the consumer AI space, but to lead it. Apple has constructed a moat so powerful it has attracted the ire of governments around the world: the App Store and iOS. App Store customers demand high-quality apps that integrate with system functions. When Siri AI hits millions of devices in the fall, people will expect their favorite apps to support it, and if they don’t, people might be nudged toward first-party apps instead. I really do think Siri AI is that compelling — it erases decades of frustration people have had with their devices. Siri AI is Spotlight search for the natural-language era of computing. However badly OpenAI wants it to, ChatGPT can never build a moat as formidable as Apple’s. Apple’s bespoke advantage is not only its impeccable product direction, but its grip on the app developer market.

Like any generative artificial intelligence system, Siri AI makes mistakes, has limitations, and often performs worse than its competitors. I just can’t dismiss these as beta bugs — they’re deeply rooted in Siri AI’s models, architecture, and marketed capabilities. As I’ve argued many times in this section, Siri AI, for now, isn’t a replacement for the LLM chatbots people are used to, and it won’t be able to do complex work. It will never do complex work; Apple has never tried to wrangle the professional AI space. But even in its new niche — sitting alongside obscure products like OpenClaw, Manus, and Perplexity Computer — Siri AI is only months away from falling behind again, as the mainstream AI companies inevitably develop some answer to it at breakneck pace. So, if I could use only one word to describe Siri AI, I’d say, “promising.” It allays fears that Apple hasn’t been working hard enough, and it brings something useful and novel to the table. That’s a good place for Apple to be, but it’s still nascent.

I’ll be closely monitoring Siri AI’s development and App Intents adoption in the months and years ahead. I expect new model updates every few months — with or without Google’s help — and I expect Siri AI to progressively gain new capabilities over time. That’s the promise of generative AI systems: they get better and more useful through incremental model releases. If everything goes to plan, Apple has a winner on its hands. The “more personalized Siri” was worth the wait.

Siri’s new voice, ChatGPT extension, and app.

Apple Intelligence

There is significant ambiguity about what counts as Apple Intelligence and Siri AI as of this year’s operating systems. This isn’t an official answer, but I speculate that “Apple Intelligence” is Apple’s brand name for all of its generative AI products, like Visual Intelligence, Image Playground, and Genmoji, while Siri AI is just one of those products housed under Apple Intelligence. (It is unclear whether the “AI” in “Siri AI” stands for artificial intelligence, or Apple Intelligence.) The honest answer is probably that Apple itself doesn’t know, and that it hastily came up with the name “Apple Intelligence” in 2024 to tout the suite as a more private, helpful version of generative AI.

Regardless, the title of this section is certainly a misnomer, because Apple blurs the line between Siri AI and Apple Intelligence far too often in this year’s operating systems. There is a subtle fifth way to invoke Siri AI on all platforms: context menus. When you select any content in iOS, macOS, or visionOS, an Ask Siri button appears at the beginning of the context menu. Tapping Ask Siri is supposed to attach files or text to a new Siri AI chat, where you can ask questions about the content. As of Beta 3, this functionality is somewhat broken: Using Ask Siri usually does not attach files, and when it does, it sometimes attaches the wrong file. It works reliably only when text is selected in first-party apps, or in third-party ones that use native text fields. This disqualifies every code editor on the market, including Xcode. There are two possible explanations for these rough edges: (a) the system suffers from simple logic errors and will be remedied before public release, or (b) the Ask Siri field is in more places than it should be, and that will be addressed in a future beta. It’s probably some combination of the two, since I think text, images, and files in Finder should be attachable, while apps on the Home Screen of iOS, for example, shouldn’t be. This is a silly but real example.

The Ask Siri menu item on iOS.

I’m including this here only because the incessant, almost Copilot-like Ask Siri field replaces Writing Tools, one of the most disappointing features from the Apple Intelligence release two years ago. There is no longer a defined set of functions Apple Intelligence can perform in text fields, such as Rewrite, Proofread, and Summarize; instead, you must ask Siri for them. Apple’s models now handle all requests; the ChatGPT integration has been removed. In a supported text field, just open the Ask Siri field or tap Write with Siri in the QuickType bar on iOS⁴ and prompt for additions or changes. Siri will then insert the desired changes into the document. It can also summarize content or, cleverly, insert a summary into a new note — all Siri AI functionality carries over. But this system does not address my key frustration with the previous Writing Tools implementation: It neither shows changes nor lets users approve or reject them. For instance, if I ask Siri to make a change in a long document, it will not indicate where it made the change or ask me to approve it. I would have to hunt the change down manually. Vexing.

Write With Siri on iOS.

Siri AI, in the current beta, immediately inserts changes into the document, which is both a regression and an improvement. In most apps, Writing Tools would display the entire edited text in a pop-up, which was next to useless. However, I do think the new functionality is presumptuous — there should be a way to approve or reject changes before Siri inserts them into the document. I am glad that there is consistency across apps: The experience is the same regardless of whether you’re working in a first- or third-party app. Previously, Writing Tools would use an error-prone, diff-like interface to review changes only in some first-party apps, TextEdit being the only one where I could reliably trigger it. Other apps would use the pop-up, and neither was a particularly useful implementation.

Write With Siri on macOS.

The replacement for “Proofread” is, mercifully, much more elegant, though still constrained. The new feature replaces the system grammar checker on both iOS and macOS, though I forgive those who don’t remember there was already one built into the operating systems. The grammar checker has always worked automatically alongside spellcheck and, on iOS, is integrated with the autocorrection system. The new feature works much the same way on macOS, underlining grammatically incorrect phrases as you type and displaying suggested replacements in the Control-click context menu. On iOS, there’s now a screen that lists all suggestions the system has made, with options to accept or ignore individual ones. This addresses my biggest gripe with Writing Tools, as explained earlier. To access it, select a suggestion as on macOS, then tap the magnifying glass icon in the Select menu. I haven’t a clue why this extremely convenient view — which even displays the reason for the correction (“incorrect apostrophe”) — is unavailable on macOS.

The new feature is slower than the previous algorithm. Corrections usually take a few seconds to about half a minute to appear, whereas the previous grammar checker was almost instant. Sometimes, the system is unusually sluggish and requires a manual trigger in Edit → Spelling and Grammar → Check Document Now. (It isn’t currently possible to trigger it manually on iOS.) Anecdotally, I don’t find the suggestions from the new system to be significantly more advanced than those of the prior checker. It still mostly corrects verb agreement, tense, and repeated words, but pales in comparison to even the free version of Grammarly, which I begrudgingly run on my Mac to catch embarrassing mistakes. Grammarly is also much faster, since it uses a complex algorithm rather than an LLM for most suggestions. For more precise editing — where intent and writing style must be intricately understood — the new system doesn’t even try, and Siri AI’s models are not large enough to act as a standalone editor; LLM chatbots are better suited for such edits. Regretfully, I cannot recommend this new system over Grammarly — it is too limited and slow to be useful.

Proofread on iOS.

These three features together — attaching files and text as context, asking Siri to edit or generate new content, and the new grammar checker — comprise the terribly named Write with Siri suite. Write with Siri is easily the most unintelligible group of features introduced in this year’s operating systems, and it deserves a lot of improvement. To briefly summarize:

There is a new Ask Siri field in every context menu in all operating systems, sans watchOS, of course. Selecting text, an image, or a file should attach it to a new Siri chat, where you can ask questions about the content. All other Siri features, like App Intents and the personal context, remain available in this mode.
Siri can also edit text if it is opened in a supported text field — just open the Ask Siri field by selecting text to attach as context, or open the field from a new line. If no content is selected, Siri can still edit the document and add new content. On iOS, iPadOS, and visionOS, a new Write with Siri button sits pinned to the QuickType bar and emulates this functionality.
The “Proofread” feature from Writing Tools is now built into the system grammar checker — suggestions are made automatically. Only on iOS, iPadOS, and visionOS is there a diff-style interface to view all suggestions in a document and approve or reject them: select a suggestion and tap the magnifying glass icon. This interface is not available on macOS.

If this is the year of polish for Apple’s operating systems, Write with Siri leaves a lot to be desired. I do think this year’s features are better conceived than Writing Tools, but that’s such a low bar to clear. It makes sense to integrate them systemwide, but the result is a patchwork of ways to prompt Siri. That itself isn’t bad, but I think the offerings are more convoluted this year, and that most people will either be frustrated by the Ask Siri prompts littered throughout the system or simply confused when the operating systems ship in the fall.

Visual Intelligence, unlike Writing Tools, has not been completely axed from the operating systems — but it now does business as Siri AI. Previously, Visual Intelligence was available in the screenshot menu and as a camera mode invoked by pressing and holding Camera Control on iPhone 16 and later. The former, introduced last year, still exists on iOS with a nearly identical interface, presenting two buttons: Ask and Search. Search is unchanged and performs a reverse image search on the screenshot; Ask, which previously used ChatGPT, now attaches the screenshot to a new Siri AI chat. This is akin to the infamous Ask Siri text field covered earlier in this section. I seldom used this feature in iOS 26 because it was slow and the ChatGPT integration was anemic, and I still don’t use it because it’s only a more roundabout way to trigger Siri’s on-screen awareness. I instead find myself opening Siri AI directly from the Dynamic Island and asking about what’s on-screen, rather than using Visual Intelligence. But I understand the Ask button is there for consistency.

Visual Intelligence through the screenshot menu on iOS.

Siri also brings Visual Intelligence to the Mac, but, just like on iOS, Apple brands it as Siri in the Screenshot app. Screenshot is an app on macOS, typically invoked using the keyboard shortcut Shift-Command-5. There is a new Ask Siri option in the screenshot menu, which takes a screenshot of a selected area and attaches it to Siri automatically; it can also be opened by pressing Shift-Command-6. If you use the keyboard shortcut — which highlights the selected area in an eye-watering, high-dynamic-range rainbow glow — there is also an Image Search button, again like iOS. I use CleanShot X for my screenshots, so I haven’t used this feature much, but I also haven’t missed it, since Siri AI on the Mac has the same automatic on-screen awareness as on iOS. I’d rather trigger Siri AI and prompt it than open Screenshot to do the same thing.

Visual Intelligence on macOS.

The camera mode also eschews the Visual Intelligence brand name. I think it’s disingenuous to call these Visual Intelligence features, but that’s what Apple calls them on its website. My working theory is that “Visual Intelligence” refers to Siri AI’s vision capabilities, but I wouldn’t be surprised if Apple phases out this branding entirely next year. It makes no practical difference for people, since the “Visual Intelligence” label is not used once in the operating systems, only in marketing. Siri is now a literal mode in the Camera app, listed alongside Photo, Video, Spatial, etc., and can still be opened by pressing and holding Camera Control. The familiar Ask and Search buttons flank the shutter button. Snapping a photo immediately sends it to Siri, whereas the Ask button takes a photo, then allows additional prompting before sending the image in a new Siri chat, similar to the screenshot menu. Images sent via Visual Intelligence are always saved to the Siri app for later viewing, and you can ask follow-up questions within the Camera app. This mode is, amusingly, also available on Apple Vision Pro, and I’ll admit it feels like something out of a science fiction movie when I use it. I think this mode is — at least on iOS — the best way to attach images to Siri AI, but I find the models worse at image processing than traditional chatbots.

The Siri mode also introduces what Apple calls “smart actions”: enhanced previews, for example, to view the nutritional information of a dish or add a custom pass to the Wallet app. My only gripe with these is that I think there’s a lot of untapped potential here for personal software. The feature is limited to only certain actions, but when they work, they work almost uncannily well — Siri could detect what food was in an image with astonishing accuracy because it’s LLM-powered. But I don’t see why these must be limited to Apple-provided use cases. I’d love, for instance, the ability to see how much sunshine and water a plant needs by just snapping a photo of it. Perhaps a price-checker would be handy, with options to view the price history of an item on select sites. People are writing (“vibe-coding”) apps like these with generative AI all the time, so an interesting way to leverage these models as they get smarter would be to have them create custom smart actions on their own. Smart actions themselves are just pleasant user interfaces around a response the model provides — I think Apple could have some fun with them.

The Siri mode in the Camera app, and smart actions.

Apple is already dabbling with personal software in these operating systems. There are two such features, both fascinating: Describe a Shortcut and Describe Extension.

Describe a Shortcut works, naturally, in the Shortcuts app. There’s a new Apple Intelligence tab in the Shortcuts editor with a simple text box. The editor now defaults to this tab instead of launching into actions, but this can be changed in Settings. Tell Apple Intelligence what the shortcut should do, and it will churn for a few seconds and return with a hopefully helpful automation. The feature is well suited for simple shortcuts, which really should be the primary objective — things like sending text messages or setting reminders, which most people would probably like to automate. Shortcuts has always been one of those oddball Apple apps that’s so helpful yet so daunting to novices, and Describe a Shortcut eliminates the barrier to entry. It does fall apart with more complex actions like AppleScript or loops — the model is trained on Shortcuts actions, not code, and probably wouldn’t be able to build something like Federico Viticci’s famous Apple Frames shortcut, which uses hundreds of actions strung together to frame screenshots. Viticci, the editor in chief of MacStories, instead made a command line utility that uses better models and can generate more complex shortcuts, and I think that’s better suited for professional Shortcuts work.

Describe a Shortcut also summarizes existing shortcuts.

But the largest barrier of all is computer code, such as JavaScript, HTML, CSS, and Swift — historically the only ways to write Safari extensions to customize websites. Describe Extension is Apple’s first foray into agentic programming with its own models, and I think the feature is transfixing in action. At its core, Describe Extension can modify websites however you’d like through natural language. Prompts like “Remove the search bar from this page” work perfectly. I’ve always injected my own handmade CSS into certain sites — infamously X — through Jeff Johnson’s excellent extension StopTheMadness to remove specific elements, but Describe Extension obviates the need for such code. Custom extensions sync automatically across devices, too. The agent can also create new content on a page, like a timer or custom UI. When you write a prompt, the system will, peculiarly, recommend existing extensions on the App Store that may already work. If you continue, Apple Intelligence will take a few seconds to generate the extension. You can then try it out and save it, edit it, or discard it. The code the model writes is not viewable.⁵

This feature is so delightful. It is virtually limitless: I asked it to “generate an extension where, when activated, it copies selected text as Markdown and encloses it in a block quotation.” I do this all the time for posts on this website. The model churned for a few seconds and came back with a nearly functioning extension: When I selected some text on a webpage and clicked the icon in the toolbar, it did almost exactly what it was supposed to, with the slight hitch that it didn’t handle new line characters correctly. I then told it about the oversight, and a few seconds later, the extension worked like a charm. Much like Siri AI itself, this feature is just so whimsical and a joy to use. It even chose an appropriate clipboard icon for the extension by itself, so it doesn’t look out of place in the Safari toolbar. You can confine an extension to a single site by asking the agent — otherwise, they’ll work on all sites. It’s aware of when the extension should only apply to a particular site, such as if asked for a specific enough change. Customizing websites is now only a few keystrokes away in Safari.

Describing an extension on iOS.

The Apple Intelligence suite this year is fortuitously much stronger than two years ago. But I still feel that most of the features were a bit rushed — I get the sense that Apple wanted to integrate Siri throughout the system but aimed too ambitiously. Ask Siri prompts litter every context menu and software keyboard in the operating systems. The new Proofread feature had real potential to compete with Grammarly — my archnemesis, and that of many other writers — but it isn’t even consistent across platforms. The Siri mode in the Camera app is clever but limited in ways that don’t make sense in the age of personal software, which Apple astutely and ingeniously embraced in Shortcuts and Safari. Don’t mistake this criticism for apathy, but I think the overall Apple Intelligence suite is still underwhelming and behind where it should be. It’s gratifying and innovative where I wasn’t expecting it to be — like Describe Extension — but underwhelms where it shouldn’t. It’s a mixed bag for sure, and it’s a real head-scratcher that Apple didn’t invest more time into getting it right, as it did with Siri AI.

Design

As I wrote at the beginning of this article, the first real feature announcement at this year’s keynote was a slider to adjust the opacity of Liquid Glass. The slider marked the beginning of what I’d call Apple’s Liquid Glass apology tour, with a litany of small tweaks and refinements to last year’s bold redesign. Liquid Glass was interesting and controversial from the start: I liked the material and the overall refresh of common controls, but I eventually admitted late in the beta process that the redesign had been too ambitious and buggy for final release. Tab bars in iOS 26 were simply illegible; toolbars blended in too much with content underneath; sidebars in macOS 26 Tahoe were atrocious; and the operating systems reeked of incompetence, with tacky icons accompanying menu items and inconsistent window corner radii. It appeared the designers of the 26 operating systems had spent too much time musing over their glass trinkets to care about the user experience.

After the perfunctory December departure of Alan Dye, Apple’s software design chief responsible for Liquid Glass, it appears saner heads have prevailed at Apple Park once again. The new operating systems look as if Apple put care, thought, and time into them. They’re not a return to the pre-OS X 10.9 Mavericks heyday of software design, but they get closer than I predicted. I’m pleased that Apple didn’t just timidly dial back the rotten parts of the 2025 redesign — it scuttled the designs that weren’t working and returned to form in many interesting ways, so much so that I had to dedicate a section of this article to documenting what has changed. The core thesis — both mine and presumably Apple’s — is that reliability was deservedly the focus in every one of this year’s design updates.

On all platforms, Apple has now modified the eponymous material used throughout the redesign — from toolbars and tab bars to prominent buttons — to increase legibility. The default Liquid Glass appearance is more opaque, frosted, and tinted. The prior material emphasized showing content underneath — like a pane of clear glass — and relied on a diffraction layer to blur it enough to retain legibility. This was a dreadful design: While attractive, it treated toolbars and tab bars as second-class, ignoring that whatever was underneath was irrelevant to a person interacting with a toolbar. It showed too much extraneous content. The new Liquid Glass material darkens content that shines through and adds a new soft blur layer. The best way I can describe it is that the old material was like a clear window, while the new one is more like privacy glass — the blur and darkness conceal the content underneath. The diffraction layer, therefore, plays a smaller role in maintaining clarity, leading to a much more pragmatic appearance. In the keynote, Apple said Liquid Glass now “diffuses complex content behind it much more effectively.”

The new default Liquid Glass appearance across apps.

The Liquid Glass slider is a genuinely peculiar feature. By default, the slider is set to the middle position. When you slide left from center, the material gets less blurry, with the far-left selection almost eliminating the soft blur layer. When you slide right from center, the blur layer doesn’t change; the material becomes more solid and eventually becomes opaque. The left half of the slider adjusts the blur layer, and the right half adjusts the opacity. This doesn’t matter to almost anyone except the most astute UI designers and nerds, like the ones reading this article — left means clear and right means opaque — but it gleans some insight into the four distinct layers of Liquid Glass: glass, blur, tint, and diffraction. Anecdotally, I think most people in the fall will, if given the option during onboarding, choose either the default or somewhere on the right side of the spectrum, since the changes are most noticeable there. The material really does get much darker, almost reminiscent of iOS 18 and earlier. I’ve tested many different settings and prefer mine in the middle of the left section.

The Liquid Glass slider in light and dark mode.

Liquid Glass elements also feature a new border and drop shadow that distinguish them from other content. These layers hug the material around the edges, giving it an almost 3D appearance. The first time I saw the new material after updating my iPhone to the iOS 27 beta, I was struck by how similar it looked to Aqua, the design language introduced in the first version of Mac OS X by Steve Jobs, who famously noted that “one of the design goals was when you saw it you wanted to lick it.” Aqua was distinctive in how UI elements cast shadows on their backgrounds, like real drops of water. The refined Liquid Glass material in the 27 releases applies this same shadow and depth effect, and it’s axiomatic where Apple’s design team — perhaps with more liberty to look back to the Jobs era — drew inspiration. The resemblance between the early-aughts-era Aqua and the 2020s-era Liquid Glass is almost uncanny in these releases. It looks gorgeous, legible, and practical, even after dialing down the opacity and blur. I’d argue it actually looks better with minimal opacity — something I’d never thought I could say about Liquid Glass, all because of the new drop shadow and improved diffraction layer.

One of the foremost gripes with the Liquid Glass redesign was the new toolbar design — plainly, that there was no longer a distinct toolbar. This “design” was abominable on iOS because it eliminated the border between content and buttons, but it was truly accursed on the Mac. That change has been reversed in all of this year’s operating systems: The toolbar background is again nearly opaque, using the same material it always has in iOS, iPadOS, and macOS. Liquid Glass buttons sit atop the toolbar and show almost no background content — even when the slider is pushed to its leftmost position — drawing a clear line between the toolbar and the content below. Tab bars still use the Liquid Glass material, but I find their appearance far less objectionable because they’re encapsulated within Liquid Glass — the material itself distinguishes the tab bar, even more so in the 27 operating systems. The old toolbar merely floated atop the content with little to no visual separation, only a light blur. It destroyed the interface hierarchy between views and controls. The new toolbar leaves no ambiguity about whether the content at the top of the screen is part of the view itself or the toolbar.

Toolbars in iOS 27.

Toolbars in macOS Golden Gate.

The new-old toolbar design has drawn some ire online from people who seemingly don’t have a clue about user interface design; they claim it doesn’t honor the transparent nature of Liquid Glass. But that was the biggest problem with Liquid Glass: It treated everything, including elements that ought to be elevated above views, as hierarchically equal. Candidly, the people who first designed it were either uninterested in the ramifications of blasting a UI with semitransparent materials, or failed to test their design properly in all apps and use cases. I must assume it’s some amalgam of the two.

On macOS and iPadOS, specific design changes are even more noticeable and appreciated. One of my callouts in last year’s OS review and my “Can Apple Fix This in 6 Weeks?” postscript criticizing the redesign’s shortcomings was the sidebar design in macOS Tahoe and iPadOS 26. Sidebars “floated” above content, and Apple’s own Human Interface Guidelines encouraged developers to extend their content underneath the sidebar. Horizontal list views would push their content past the sidebar to the edge of the window, and the HIG said to scale up media knowing that the sidebar would partially obscure it. The sidebar design was a complete disaster, and may have been one of the worst designs Apple’s typically star-studded Human Interface team has ever cooked up. This design is now gone in macOS Golden Gate and iPadOS 27, replaced by a near-replica of the macOS 11 Big Sur-era sidebar. There is again a distinct border between the sidebar and the detail content, and the sidebar is no longer inset. It’s instantly recognizable as soon as you boot into macOS Golden Gate — it just looks and feels right.

The Music app’s sidebars in macOS Golden Gate.

All apps — including those that aren’t compiled for the new operating systems — receive these new design improvements. Third-party apps in macOS Golden Gate look great with no developer work, especially the ones that put significant care into their macOS Tahoe redesigns. The same goes for iOS and iPadOS, though some developers put extra work into making their apps look more habitable on iOS 26, and now that Apple has addressed many of the design shortcomings, those considerations are obsolete. I’d say most developers will have to put little to no work into finalizing their apps for the new releases in the fall. All apps on macOS — including those that haven’t even been redesigned for macOS Tahoe — use a new, tighter corner radius that truly makes me happy. It’s slightly more aggressive than macOS 15 Sequoia, the pre-Liquid Glass version, but a striking improvement over the bulbous, Fisher-Price-esque windows of macOS Tahoe. macOS Golden Gate just looks like it was designed by adults who know what they’re doing. Apple’s vapid obsession with “concentricity” is gone, and the design looks thoughtfully crafted.

The menu item icons are mostly gone in macOS Golden Gate.

The unified window corner radii in macOS Golden Gate.

Every aspect of the design has been given a once-over with a fine-tooth comb. The incessant menu item icons are now gone by default. Notifications now adapt better to light-colored Lock Screen wallpapers and even drop down from the top-left corner to indicate the new Notification Center gesture. Animations throughout the system are smooth as butter — I have yet to notice even a frame drop while using the first few betas. Sidebar icons on iPadOS and macOS are tinted again and even fade to gray when the app isn’t in the foreground, just like they have for decades. App icons created through Icon Composer are automatically rendered to remove the “fuzziness” that plagued them last year, with no developer intervention necessary⁶. I could go on about all of the little details that make the new operating systems feel like they’ve been crafted with care.

Liquid Glass in macOS Golden Gate, using my preferred clearer appearance.

I can’t overstate how many people have come up to me — strangers and acquaintances — in the last year to complain outright about the iOS 26 design. The criticism is harsh but not undeserved: iOS 26 by default is almost unreadable, and macOS Tahoe truly is one of the worst macOS releases in a long while. People couldn’t even read the time on the Lock Screen without switching to the Solid clock appearance, for heaven’s sake. The updated Liquid Glass design in iOS 27, iPadOS 27, and macOS Golden Gate goes a long way toward addressing the havoc of last year’s rushed releases and cherishes the fresh coat of paint I and other UI design enthusiasts praised in our reviews of the new UI paradigm. I think this will be the year most corporate third-party developers will adopt the new design, especially since they’ll be required to compile their apps with Liquid Glass beginning in the fall. It’s allaying to see Apple take developer and user feedback to heart.

Photo Editing

I have unwavering respect for my readers, so this section does not contain a fourth cathartic tangent deriding AI-generated imagery. But Apple has added some new generative AI image-editing tools across the system that are worth briefly covering.

Two years ago, Apple announced Clean Up in the Photos app, a generative AI feature for removing unwanted objects from images. This year, the Photos app improves Clean Up with new models and adds two AI photo-editing tools in an apparent effort to better compete with Google’s offerings: Extend and Reframe. The new modes are housed in the new Tools menu of the photo editor on all platforms, and all work with varying levels of success.

The enhanced Clean Up tool uses two new models: Fast and High Quality. Users can choose between the two manually, or select Auto, which recommends the best model depending on the task. The Fast model works as soon as an object is highlighted in the image, editing the photo in only a few seconds once the model loads. Its accuracy, I’ve found, is no better than that of the prior Clean Up model, and it still produces sloppy images; I wouldn’t recommend using it for much of anything, even small touch-ups. The larger, High Quality model appears to send images to the cloud and takes much longer — up to a few minutes — to make edits. When it’s selected, it asks you to mark all edits at once, then tap the Clean Up button when finished to let the model go to work, so you aren’t stuck waiting separately for each edit. The results are uncannily good — the larger model is worth the wait, especially with the batch-processing interface.

The improved Clean Up feature in iOS 27.

The Extend and Reframe features are much less useful. Extend expands the photo around all four edges, using what’s already in the photo to craft a realistic scene. When you first enter Extend, it initially displays the original photo with blurred, “dreamy” edges. You can then expand the photo using the drag handles, or by pinching to zoom out. I’ve found that the feature can generate a new image up to double the size of the initial one, but it isn’t possible to limit Extend to only one side of the image. Users cannot, for instance, strictly extend the top edge without altering the other three sides to some degree⁷. After tapping the Extend button, it takes a few minutes to generate the new photo. It’s far riskier than Clean Up, especially because it works with every image, including highly complex ones where it would be difficult, if not impossible, to predict what was outside the frame. Extend will, for instance, give pants to someone wearing shorts or create people who don’t exist in the background. Every photo modified with it clearly looks AI-generated — nobody will be fooled. Extend also replaces the Extend Background feature on the Lock Screen for ill-fitting photos, and I think that’s perhaps the only place it’s suitable.

The Extend feature in iOS 27.

Reframe works much the same way, but I think it has even less utility. It’s effectively the same feature as Extend — presumably using the same models in the cloud — but it also alters the Z-axis, or depth. It first takes a few seconds to create a depth map of the image, then lets you pan around using the same “blurry edges” interface from Extend. After a few minutes, it “fills in the gaps,” as Apple says, generating the missing edges to “reframe” the photo. It, too, is unconvincing to even a casual viewer. Images clearly look manipulated at best, or completely fake when the model struggles, and the degree of movement is quite limited. The models are nowhere near as capable as ChatGPT or Google’s Nano Banana image generator, built into Gemini. And I don’t think it’s anywhere near as useful as some of Google’s or Samsung’s AI image gimmicks, which, while sloppy, at least have some real-world value. Reframe and Extend don’t seem to me like very intentional or well-crafted features, and I don’t think they should’ve shipped in their current state, if at all.

The Reframe feature in iOS 27.

Image Playground, Apple’s app for creating AI-generated images, is much the same, though it adds a “photorealistic” style that really does look terrible. Perhaps it’s due to Image Playground’s limited availability, but I haven’t heard of a single person using it in the real world, whereas I do know image generators are rife among office workers for presentations and fliers. Free tools like Nano Banana and ChatGPT work better, so I keep asking myself why Apple intends to be in this market. Do Apple employees use it often internally? That doesn’t seem aligned with the company’s culture at all, so maybe a handful of executives who initially came up with the idea two years ago are intransigent about letting it go. I can’t see the new model being any more popular than the last one, given that even when Image Playground integrated ChatGPT last year, it didn’t seem to gain any more users. Image Playground ought to be shown the door.

Feelings aside, I just don’t think Apple is suited to deliver generative AI image features. Image generation is complicated — even for the major AI labs spending hundreds of billions of dollars on pre-training — and Apple’s models are not powerful enough to ship on the most popular smartphone in America. People will stumble upon these features while they’re editing photos on their new iPhones in the fall, but they’ll either find them useless or — worryingly — use them anyway, subjecting everyone to a flood of deceptive, poorly made content online. Apple felt internal and external pressure to ship image generation in 2024, but I don’t believe that pressure exists anymore. Clean Up is the sole exception to this: It was fine in 2024, and it’s good enough now. I don’t necessarily have a problem with people using it. But Extend, Reframe, and Image Playground don’t meet Apple’s standards, and their premise should be rethought entirely.

Miscellany

I began this review over 11,000 words ago with an intrepid claim: that this year’s operating systems hearkened back to the age of reliable yet innovative Apple software. Indeed, iOS 27 and macOS Golden Gate are the most reliable Apple platforms since iOS 12 and macOS 10.14 Mojave, when the company placed great emphasis on speed, performance, and bug fixes. Apple displayed a slide with close to 100 optimizations made this year, including a new Spotlight index, faster app launch times on older devices, and improved connectivity switching on iOS. I can’t remember the last time — if ever — the company spent so long describing changes to the CPU scheduler.

Everything in these releases feels faster and less buggy, even on my iPhone 17 Pro and relatively new M3 Max MacBook Pro from 2023. Everyone will feel the improvements, not just those using older devices. I’d argue they make even more of a difference than Siri AI and most of the other features introduced in the past few years. Some changes I’ve noticed include:

Faster and more reliable search in Mail, a feature that has languished for decades.
Correct results when searching System Settings, which has been partially broken since iOS 16 and macOS 13 Ventura.
Improved AirPods automatic device switching, which was too eager to switch to a Mac when AirPods were connected to an iPhone.
Faster and more reliable iPhone Mirroring.
Much better battery life — up to two hours more on iOS per charge and an hour more on my degraded MacBook Pro’s battery.
Much smoother performance while using Low Power Mode; it was almost unusable in iOS 26.
More honest Screen Time metrics on iOS and macOS, and improved syncing across devices.
More reliable automatic switching to cellular data when a Wi-Fi network is down; there is also a new Connectivity Assist, replacing Wi-Fi Assist, which worked sporadically.
System apps like Passwords now remember the sidebar size across launches on macOS.

There are some additional features I didn’t have space to cover elsewhere in this review.

The Passwords app will use an agentic computer use model to change compromised passwords on a user’s behalf. I don’t have any weak or compromised passwords to change — it’s only available if you do under Passwords → Security — but I am eager to test this out. This is one of Apple’s most ambitious generative AI features.
Keyboard dictation on iOS can now use the new, reliable dictation model that debuted last year in Notes and Voice Memos. I hoped for this in last year’s review, and it works like a charm. It’s only available for iPhone 17 and later models and iPads with M-series processors. Apple’s website indicates it should work on Macs with M3 or later chips, too, but it isn’t yet displayed under System Settings → Keyboard → Dictation → Advanced Dictation Preview, as of macOS Golden Gate Developer Beta 3.
Apple Intelligence now analyzes and summarizes HomeKit Secure Video footage in the Home app, and a new search bar finds relevant clips with natural language. I’ve found this feature hit-or-miss. Clips have a new appearance in the camera view, and there is a new unified All Cameras page with a list of clips from all video cameras in a home and a summary next to each clip. Perhaps it’s because my cameras are not 4K, but I’ve found the summaries unhelpful. Most are “Motion was detected,” “Someone was detected,” or, blankly, “A summary is not available for this video.” It seldom detects when packages were dropped off or if animals were spotted — I’ve never had this problem in previous versions.
The ChatGPT extension in Siri, released in 2024, hasn’t been removed. If enabled under Settings → Siri → ChatGPT, you can choose to send a prompt to ChatGPT in both the Siri app and Spotlight. (Voice queries are not routed to ChatGPT, and you must select ChatGPT every time; you cannot set it as the default assistant.) ChatGPT does not have access to the personal context and cannot call App Intents, which makes the Siri AI system unavailable in the European Union.
Messages and Mail now offer smart suggestions powered by Apple Intelligence, based on the message content. These suggestions allow people to add, say, a reminder in Reminders or a note in Apple Notes about the message when relevant. They work reliably, though I think it’s a shame that third-party apps can’t offer to be suggested apps or show suggestions themselves.
The Settings pane for connected AirPods has been redesigned. It also now includes a new equalizer, or EQ, selector to adjust the volume of mids, highs, and lows. The selector doesn’t significantly alter the sound profile, since AirPods already sound fantastic, but they tend to be mid-heavy, which doesn’t favor genres like hip-hop or rhythm and blues. I can see this being popular in the fall.
Nearly five years after the notch was first added to the MacBook Pro, macOS Golden Gate hides menu bar applets behind a double-chevron icon when the notch would otherwise obscure them. Without a menu bar manager app like Bartender or Vanilla, menu bar items that didn’t fit in the limited space to the right of the notch would disappear. This worsened when an app’s menu items overflowed to the right, further limiting the space for applets. If either of these scenarios occurs in macOS Golden Gate, macOS will instead automatically hide the overflowing items behind a new icon. (This temporarily breaks menu bar managers during the beta.) I cannot believe such a feature took so long — do Apple employees not use menu bar apps?

The menu bar chevron in macOS Golden Gate. (The blank space is the notch.)

macOS now keeps a more comprehensive list of apps that run in the background under System Settings → General → Login Items & Extensions. This list was previously limited to daemons in the ~/Library/LaunchAgents and ~/Library/LaunchDaemons folders; it now lists apps registered under the background task management database introduced in macOS Ventura⁸. From this list, it is possible only to disable these apps from running in the background. To remove them, such as after uninstalling the app, run sudo sfltool dumpbtm in Terminal, find the file paths of the apps you wish to remove, and rm them.

The background task management database in System Settings.

Apple Intelligence in Safari can automatically create special tab groups called topics. It works well enough.
Genmoji, the feature that allows people to create custom emoji stickers, has been updated with a new image generation model. The results are almost identical to those of previous versions, but the model is much easier to prompt. It adheres to instructions more closely, uses fewer words, and better understands intent. (I still don’t know of a single third-party app that supports Genmoji, despite the billboards plastered throughout Manhattan.) Genmoji is also fully available on macOS Golden Gate through the Messages app.
Image Playground can now generate wallpapers and is presented as an option alongside Photos, People, Color, etc., in the wallpaper chooser. It works just as well as anyone would expect, which is to say not well.

Image Playground wallpapers, new AirPods settings, and new Mail search.

This year’s review came in at nearly 4,900 words shorter than last year’s. For that reason, I don’t think Apple is likely to replicate this year’s releases for a while, probably half a decade at minimum. WWDC is a press event designed to garner attention — the company wants people and the press to get excited about how their existing iPhones will become even more capable and feature-rich in just a few months, all for free. And most years, that strategy does pay off. The contingent that resents iOS updates has slowly dwindled since iOS 7; almost everyone updates to the new version of iOS by spring. But I think Apple knows that iOS 26, while it had nearly identical adoption numbers on paper, was quietly resented. It did too much at the expense of usability.

I said this at the beginning, almost 13,000 words ago, and I’ll repeat myself: iOS 27 is a return to form for a company that just a few months ago felt like it had lost its way. Marred by politics — internal and external — regulatory restrictions, bad press, C-suite changes, and a heavily criticized two-year delay of generative AI features that were already late when they were announced, Apple was not in the best spot earlier this year. I think iOS and iPadOS 27, macOS Golden Gate, and Siri AI put the company in a much better position, building trust with the press, developers, and probably users. People will find Siri AI genuinely useful in all sorts of scenarios, from searching within apps to automating odd tasks. Each of the new platforms was designed with care and respect for the user experience, qualities unseen from Apple’s Human Interface team for far too long. There are missteps, but overall, this year is a great one for Apple software. I don’t say that every year, but I wish I did.

iOS 27, iPadOS 27, macOS Golden Gate, and Siri AI are available in public beta beginning Monday.

Siri’s new “expressive” voice is only available on certain devices. My preliminary impressions are that it’s not a significant improvement over the old voice, and it remains miles behind ChatGPT’s voice mode or Gemini Live. I’m not perturbed, though, since I never turn on the Speak Back for Voice Input option. ↩︎
Anecdotally, I find Siri AI’s hallucination rate is lower than that of GPT-5.5 Instant, OpenAI’s default model, though this could be because I haven’t been asking it tough questions. Siri AI does a fantastic job of providing links to external sources and often chooses to quote them rather than summarize or rewrite — an interesting and manifestly intentional design choice. As of Developer Beta 3, Siri AI also writes a disclaimer at the bottom of each response: “By the way, always verify important details. As an AI, I may make mistakes.” The disclaimer is written in plain text as part of the response, but it’s always reliably there. I wonder if Apple will move it into the interface instead, as some other chatbots have, or if it will conspicuously remain in the response. ↩︎
One of the only new features in Apple Notes this year is that it’s now possible to paste Markdown into Notes and have it automatically converted to rich text. My hunch is that Apple added this primarily for Siri AI, since it would’ve been harder and less extensible to train the model to write rich text over Markdown, the de facto language of every AI model. ↩︎
This must be a beta bug — it genuinely feels like something Microsoft would do. (Cupertino, start your photocopiers.) Simply put, you can either select text to provide it to Siri as context, or select no text at all and still invoke Siri. This works across platforms. The button in the QuickType bar, limited only to iOS, iPadOS, and visionOS, is redundant and does the same thing as selecting no text. ↩︎
It is viewable during the developer beta period, but you must submit a bug report to read the code. Feedback Assistant attaches it to the report automatically. ↩︎
App icons comprise three layers beginning in the 26 operating systems: the background, the glyphs, and Liquid Glass. Developers provide these elements individually to the system, which renders them on-device. The previous rendering algorithm didn’t properly distinguish the glyph from the background, making the glyph’s edges appear soft and blurry. In the new operating systems, edges look crisp and polished. The Liquid Glass effect on top has also been made less aggressive, and there’s an optional clear Liquid Glass material within icons, used by the Maps app. ↩︎
The blurry edges in Extend mode subtly indicate that it will always modify all four edges; I think Apple did a phenomenal job designing the UI. It is possible to request that there be more content on one side, but all four sides will be modified to some degree anyway. It’s slightly unintuitive, but I think the company did its best given its models’ capabilities. ↩︎
Apple’s support documentation disputes that BTM-registered apps weren’t shown in this list, but I think I’m correct. I indeed see many more apps in this list after updating to macOS Golden Gate, none of which are listed in the aforementioned Library subdirectories. They’re registered in the BTM database. My guess is that enterprise-configured Macs — which the support documentation is written for — did show all BTM-registered apps in this System Settings page, but non-enterprise Macs didn’t until macOS Golden Gate. ↩︎

Apple Is Suing OpenAI and Employees for Stealing Trade Secrets

Eshu Marneedi — Fri, 10 Jul 2026 17:53:10 -0400

Chance Miller, reporting for 9to5Mac:

Apple has filed a lawsuit against OpenAI today, accusing the company of trade secret theft. Specifically, Apple alleges that its former employees have stolen trade secrets “for the benefit of OpenAI.”

“This case is about Apple’s former employees stealing Apple’s trade secrets for the benefit of OpenAI. Apple brings this suit to put a stop to it,” the lawsuit says…

The complaint, filed in the U.S. District Court for the Northern District of California, alleges that Tan used insider knowledge of Apple’s confidential projects to grill job candidates in interviews. Additionally, Tan directed job candidates still working at Apple to bring actual Apple hardware components and samples for “show and tell” sessions.

Furthermore, Apple says a candidate began “screenshotting and downloading files relating to a highly confidential Apple project” hours before interviewing with Tan, who then “solicited more information about that same Apple project” once the interview started. This became an “established pattern,” Apple says.

Tan also allegedly possessed and distributed an internal Apple “Need to Know” document to new OpenAI hires before they gave their notice to Apple. The document included Apple’s departure security protocols. As part of its investigation, Apple found a “pattern by employees who depart for OpenAI of taking steps to evade the security processes intended to protect Apple’s confidential information.”

I read the full complaint. Some notes:

Chang Liu, an ex-Apple employee who now works for OpenAI, allegedly (a) failed to return all of his company-provided devices before he left, (b) used an exploit to access Apple’s internal files, and (c) downloaded them while working for OpenAI.
Tang Tan, OpenAI’s chief hardware officer — and Jony Ive’s right-hand man at OpenAI’s “io” hardware division — allegedly tells potential employees during interviews to bring in unreleased Apple hardware for “show and tell.” And as Miller notes, Liu and other ex-Apple employees suggested new hires “prepare” for their OpenAI interviews by studying unreleased Apple products.
Tang and Liu only got caught after they used work-provided devices to — in some cases — brag about their actions.
Over 400 ex-Apple employees work at OpenAI.
OpenAI allegedly contacted an Apple hardware supplier and asked them to “perform Apple’s proprietary, trade secret processes for OpenAI’s benefit.”

I’m not really surprised at any of this. There seems to be bad blood between the two companies — OpenAI’s models weren’t spoken about once during the Worldwide Developers Conference State of the Union in June — and OpenAI had been loosely threatening legal action even before WWDC. Moreover, OpenAI has never come across as a particularly honest corporation. Sam Altman, its chief executive, is a seasoned liar and manipulator; OpenAI employees are known for being narcissistic online; and the company doesn’t seem to attract very honest, forthcoming fans or workers.

But it does seem illuminating that OpenAI has such an inflated image of itself that it thought it could successfully evade one of the most litigious and secretive companies in the world. Apple does not let such behavior go unpunished — it never has and it never will. Everyone knows not to mess with Apple Global Security, except, clearly, OpenAI. The OpenAI people clearly have become accustomed to a privileged position where they feel untouchable. They were handed an inconceivable amount of free cash and thought it could protect them from anything, including immoral and illegal behavior. Maybe it will, but it doesn’t seem likely.

The only solution to OpenAI’s immoral behavior — and there are so many instances of it — is to force them to start turning a profit. It will happen eventually; the investors will get tired of losing money, and they’ll put pressure on the company. Perhaps this lawsuit is a step closer to that reality.

OpenAI Merges ChatGPT and Codex Apps, for Apparently No Reason

Eshu Marneedi — Thu, 09 Jul 2026 18:13:23 -0400

OpenAI, on X:

Introducing ChatGPT Work, a new agent in ChatGPT powered by Codex and GPT-5.6. It can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work. It’s a whole new way to get work done.

ChatGPT Work is powered by GPT-5.6. GPT-5.6 makes ChatGPT state of the art at reasoning through complex tasks and creating materials that match your templates, reference files, and preferred style. Just describe the outcome you want, without having to spell out every step to get there.

On web and mobile, ChatGPT Work is rolling out today for Pro, Enterprise, and Edu plans. It will roll out to Plus and Business plans over the next few days. In the ChatGPT desktop app, Chat, Work, and Codex are available on every plan, including Free, and is available globally to download on Windows and Mac.

Codex app users can update their app as usual — it will become the new ChatGPT desktop app.

I fear OpenAI’s employees have lost the plot. The old, native ChatGPT app for macOS still exists and is renamed “ChatGPT Classic” after updating to the latest version. The Codex app, after updating, is renamed “ChatGPT.” The unfortunate side effect of this is that if you update the Codex app before the ChatGPT Classic app, the two names will conflict, and macOS will refuse to open the Codex app. This happened to me; I had to reinstall the Codex app from the OpenAI website to fix the issue.

Now I have two ChatGPT apps on my Mac: ChatGPT Classic and ChatGPT, née Codex. The ChatGPT Classic app has languished for years now, but when it was announced, it truly was incredible. It’s beautifully crafted with care, down to the little details. There’s even a mode to let ChatGPT peek into your apps like BBEdit or TextEdit — this is how we used to code with artificial intelligence before Codex and Claude Code! The Codex app is made with Electron, but I’ve always regarded it as one of the best Electron apps on the Mac. It’s not perfect, but superior to Claude in every dimension.

The new update ruins months of good favor with the Codex app. It quite literally benefits nobody. There are now two primary modes in the ChatGPT app: Codex and ChatGPT Work. Codex is the same; Work is just a simpler version of Codex, similar to Claude Cowork. I don’t even know why this mode exists — Codex has always been simple enough for non-technical people to use anyway. It’s also redundant in the cloud (iOS and the web) because it’s only limited to using plugins, which the standard ChatGPT mode can do as well. There’s almost no end-user benefit to using ChatGPT Work on the web or on iOS.

The actual ChatGPT interface — what the app is named for! — is hidden in a mini-window accessed from the sidebar. This is the most absurd, counterintuitive user interface OpenAI has ever conceived. If the app is literally named “ChatGPT,” how is ChatGPT not the main interface? Does OpenAI see ChatGPT Work as the replacement for ChatGPT, and if so, why doesn’t it make that clear? And to make matters worse, there’s no consistency in the models across interfaces. Work and Codex have GPT-5.6 Sol, Luna, and Terra — the three sizes of GPT-5.6 — while ChatGPT only exposes the Instant, Medium, and High reasoning levels. Medium and High are GPT-5.6 Sol, though that is not indicated in the interface.

The ChatGPT app doesn’t open ChatGPT anymore, and the models once again make no sense. Here’s a handy chart of the models and their reasoning options across ChatGPT modalities:

	GPT-5.6 Luna	GPT-5.6 Terra	GPT-5.6 Sol	GPT-5.6 Sol Pro
Work and Codex	Light, Medium, High, Extra High, Max, Ultra	Light, Medium, High, Extra High, Max, Ultra	Light, Medium, High, Extra High, Max, Ultra	N/A
Chat	N/A	N/A	Medium, High, Extra High	For Pro users only.

I haven’t heard from a single person happy with Thursday’s changes. Codex users are upset because their app has been renamed; ChatGPT users on the Mac are upset because the new ChatGPT app really isn’t ChatGPT at all. The only people seemingly pleased with this “work” are OpenAI employees themselves, who seemingly don’t understand anything about marketing or communications. But I’m supposed to believe this company is worth a trillion dollars.

A correction was made on July 13, 2026, at 5:30 a.m.: Previous versions of this post stated that GPT-5.6 Terra and GPT-5.6 Luna were available in ChatGPT Chat. This is incorrect; these models are limited to ChatGPT Work and ChatGPT Codex. Free users — and paid ones who select the Instant reasoning effort — are defaulted to GPT-5.5 Instant in ChatGPT Chat. In ChatGPT Work and Codex, free users are limited to GPT-5.6 Terra and GPT-5.6 Luna. Plus and Pro users can use GPT-5.6 Sol in ChatGPT Work, ChatGPT Chat, and Codex. Chart:

	GPT-5.6 Luna	GPT-5.6 Terra	GPT-5.6 Sol
Free and Go	Available in Work/Codex.	Available in Work/Codex.	Unavailable.
Plus and Pro	Available in Work/Codex.	Available in Work/Codex.	Available in Work/Codex/Chat.

Zuckerberg ‘Admits’ Meta’s Layoffs Were Ineffective

Eshu Marneedi — Fri, 03 Jul 2026 01:02:30 -0400

Mark Zuckerberg, Meta’s chief executive, at a Meta town hall reported by Katie Paul and Courtney Rozen at Reuters:

In retrospect, he said, the “trajectory of the agentic development over at least the last four months hasn’t really accelerated in the way that we expected,” and that the company’s bets on the new structure “haven’t come to fruition yet.” Zuckerberg was referring to AI agents, automated systems that can execute tasks on behalf of a user.

Conversations he was having “with our top people” when they started planning the restructuring in January and February “were that they were worried that we weren’t going to move fast enough to adapt,” Zuckerberg said.

At the time, he said, executives were “super optimistic” about tools like Claude Code from AI startup Anthropic.

The self-created tragedy of Meta is that the company is loath to invent new products. Instead, Meta’s management more or less relies on “vibes” to govern its decisions, and those vibes are often either wrong or far too late. The most pertinent example of the former is the ill-fated metaverse, which was developed solely on the (unbelievable) whim that the pandemic would last forever — or at least far longer than it actually did — and people would become accustomed to replacing in-person interaction with virtual reality. It was precisely at this moment, roughly around mid-2020, that Meta (then Facebook) disintegrated from a social media company into a Ship of Theseus that still technically operated its core social platforms but fundamentally was distracted by a red herring. Vibes-based management.

As I wrote in my now infamous “Meta-stasizing Cancer of Indirection” piece, Zuckerberg did not learn from this disastrous failure as the artificial intelligence boom kicked off in 2023. Long story short: Zuckerberg threw his company into turmoil because he was too late to identify that the metaverse was an abysmal failure. By the time he did, the AI boom was already in full swing, and Meta was thoroughly left out. This strategic failure, coupled with Zuckerberg’s arguably incompetent management style, left employees either out of employment, directionless, or both. It is just impossible to run a company on a whim — the metaverse was a distraction, and so was AI because Meta was far too late and improperly organized. Vibes-based management.

Zuckerberg yet again plunged his company into chaos after the success of Claude Code in December 2025. Knowing the company was behind in developing AI products after observing the rise of agentic coding, Zuckerberg effectively put Alexandr Wang, the chief of Meta’s AI division, in charge of the entire company. The only thing Wang did was wrongly determine that all human programmers were a waste of time and money and that it would be better to fire them and spend the freed-up cash on talented AI engineers who would unwittingly develop their own replacements. So that’s exactly what Zuckerberg did, per Wang’s hunch: he fired thousands of employees, put AI in charge of content moderation, and mandated that the remaining Meta workers install spyware that would track their computer use to train an agent that could take their job. Vibes-based management.

I can’t tell if Zuckerberg is dimwitted or just evil. The problem during the first era of the AI boom (circa 2023) was indeed that Meta was too slow to identify the metaverse flub. But that was no longer Meta’s problem entering the agentic coding era: The problem, rather, was that Meta had no coherent strategy. The last thing it should’ve done was “move fast enough to adopt” because “adopting” was not the answer to Meta’s problems. AI-assisted programming has developed in the last six months — contrary to Zuckerberg’s claim that it “hasn’t really accelerated” — but indeed not in the way Meta expected, because Meta’s vibes-based management this time was just plain wrong. AI never had the potential to replace so many workers at an instant. The vibe was — unlike in 2023 — not late, but wrong entirely. And I’m confident in saying only a fool could have lent credence to that laughably incorrect theory.

I’d say getting fired by Meta is like catching the last plane out of Vietnam, even in this ruthless job market.

6 Years Later, Is 5G mmWave Widespread?

Eshu Marneedi — Thu, 02 Jul 2026 01:24:01 -0400

Marko Zivkovic, reporting for AppleInsider:

For the U.S. variant of the iPhone 18 Pro, which will feature mmWave compatibility, Apple seemingly plans to use Qualcomm modem hardware.
Multiple Qualcomm components, including the SDX80M, SDR875, QDM8771, QDM8720, PMK75, PMX75, and QET7100A, are referenced in a bill of materials related to the iPhone 18 Pro model Apple plans to sell in the United States.

As for the iPhone 18 models which will be sold elsewhere, Tata documentation suggests these configurations will use Apple’s proprietary C2 modem. While this approach may sound unusual, there is at least one possible explanation.

Apple’s current in-house modems, the C1 and the C1X, do not support 5G mmWave, and it looks as though the C2 will continue this trend. Until Apple develops a modem compatible with mmWave, it looks as though the company will offer mmWave support to iPhone 18 Pro users by using Qualcomm hardware.

I think Apple made the right choice in developing the C2 modem without support for 5G millimeter wave (“mmWave”), the technology that enables ultra-high-bandwidth connections in limited areas. Millimeter-wave signals — known by carriers as ultra-wideband — are extremely high-frequency and can be interrupted by even atmospheric attenuation. They cannot permeate through walls or people and do not work across long distances. This makes them unsuitable for most cellular transmissions; carriers instead use the lower-frequency C-band to transmit slightly slower but more reliable 5G signals. (Confusingly, carriers also refer to this as “ultra-wideband.”) The three major carriers — Verizon, AT&T, and T-Mobile — have installed 5G ultra-wideband antennas in city downtowns across the United States. Some sports arenas and airports have also been outfitted with the technology to limited success.

During Apple’s “Hi, Speed” October 2020 keynote, which introduced iPhone 12 with 5G, the company incessantly touted 5G ultra-wideband as the future of mobile communication. So let’s take a step back for a minute: Was the rollout of 5G ultra-wideband all that successful? Is it really everywhere? For one, Verizon, with which Apple partnered during the keynote, has since removed the map on its website that showed where 5G millimeter-wave was available. The coverage map now only marks certain metropolitan areas as “5G ultra-wideband capable” cities, while refusing to discriminate between the much slower C-band and millimeter-wave connectivity. There is, however, a map buried on a subdomain (gismaps.verizon.com) that shows the millimeter-wave rollout, and it’s quite underwhelming.

The map shows that Verizon still has not rolled out 5G millimeter-wave on even some of the most population-dense Manhattan streets. This is not malfeasance: It’s more likely that Verizon, sometime in the last six years, determined that it was either impractical or impossible to bring 5G millimeter-wave to most streets and instead slowly phased out the advertising. Nowadays, Verizon hardly even mentions 5G because it is, even on C-band, mostly a waste of battery. I have never once connected to millimeter-wave in these six years — despite living in an “ultra-wideband capable city” — and I assume most Americans are the same. If not for New York, where would people be most likely to take advantage? 5G millimeter-wave was a profound waste of everyone’s time and money, and the carriers know it. Betteridge’s law of headlines strikes again.

What isn’t a waste of time and money, though, is Apple’s C-series modems, which have performed equally as well as the Qualcomm modems — if not a bit better — while using less energy. If the carriers, Apple, and customers all somewhat agree that 5G millimeter-wave is a profound waste of resources, why even ship it at all? Carriers don’t even make a distinction between the two flavors of ultra-wideband anymore, and almost no one who buys iPhone 18 Pro will even know the difference. How many people are standing in Times Square running internet speed tests? I think people, including the carriers, would be happy to trade a technology hardly anyone uses for an hour of extra battery life a day with the C2 modem. I think it’s a mistake that Apple hasn’t recognized this and used the C2 in all iPhone models.

Apple Raises Prices Across Almost All Products, Effective Immediately

Eshu Marneedi — Thu, 25 Jun 2026 19:18:25 -0400

Hartley Charlton, reporting for MacRumors:

After temporarily taking it down earlier today, Apple’s online store is back up with a series of product price increases. The changes are as follows:

HomePod mini: $129, up from $99 (+$30)

HomePod: $349, up from $299 (+$50)

Apple TV: $199, up from $129 (+$70)

iPad: $449, up from $349 (+$100)

iPad mini: $599, up from $499 (+$100)

iPad Air: $749, up from $599 (+$150)

iPad Pro: $1,199, up from $999 (+$200)

MacBook Neo: $699, up from $599 (+$100)

MacBook Air: $1,299, up from $1,099 (+$200)

MacBook Pro: $1,999 up from $1,699 (+$300)

iMac: $1,499, up from $1,299 (+$200)

Mac mini (M4 Pro): $1,599, up from $1,399 (+$200)

Mac Studio (M4 Max): $2,499, up from $1,999 (+$500)

Mac Studio (M3 Ultra): $5,299, up from $3,999 (+$1,300)

Vision Pro: $3,699, up from $3,499 (+$200)

I am humbled to admit I was wrong in my post earlier in June about these price increases: I predicted they wouldn’t take effect until new products were announced and that the iPhone (and Apple Watch) would be the first to take the hit. The iPhone and Apple Watch did not receive price increases on Thursday, perhaps in anticipation of the new models in September. Every other product, however, saw preemptive double-digit percent increases. I’m not willing to bet anything on it, but I feel like this indicates that most of these products — including the Apple TV and HomePod mini — won’t be updated until late in the fall, or maybe even next year. I don’t believe the (somewhat sensible, if not credible) theory that Apple is increasing prices to “get people used to” the higher price tags before new models are released later this year. (I’ll return to product cycles in a bit.)

Apple makes exceptional margins on its products, close to 40 percent excluding Services revenue. (Including Services, it makes about a 45 percent margin.) Some memory prices have increased by 100 percent year-over-year, but Apple has locked-in contracts with RAM manufacturers dating back years. It’s undeniable that despite that, Apple’s margins have taken a severe hit in recent months due to the RAM shortage — so much so that it may hinder research and development — but Thursday’s news seems to indicate that the company wants to return its margins to those industry-leading levels, not meet somewhere in the middle. I think that’s highly unrealistic, and Wall Street seems to agree, at least preliminarily: Apple’s stock dropped 6 percent after the news broke.

There are two sides to this coin if you’re Apple’s chief executive. If Apple raises prices, people will save their money and keep their existing Apple products, which work reliably for years, causing the company’s profit to tank in the upcoming quarters. But if Apple doesn’t raise prices, or raises them less heavily, its profits would also fall, and it might have to raise them even more aggressively in the future anyway just to maintain any profitability. Apple’s C-suite evidently opted for the short-term blow, hoping that people would accept the price increases in the months ahead. Wall Street, however, did not run Apple stock into the ground whenever the market rate for RAM increased (i.e., when Apple’s margins declined) — it believed, somewhat convolutedly, that maintaining prices and taking a (perhaps temporary) hit on margins would still be more profitable in the long run. It instead tanked the stock on Thursday.

I tend to agree with Wall Street here. These price increases are genuinely unconscionable. $250 for an Apple TV with Ethernet, whether it runs Apple Intelligence or not, is ludicrous. The iPad Pro is not worth $1,200. And Apple simultaneously increased prices on its refurbished products — which it does not have to pay to manufacture — while maintaining trade-in values for customers who wish to upgrade. On the surface, especially to anyone not well-versed in RAM supply-and-demand economics, these price increases look like the most blatant cash grab in Apple history. Especially for a company known worldwide for its historic profit margins. An M5 Max MacBook Pro now costs $4,100 without any memory and storage upgrades — that’s a $500 price delta overnight. Apple products, for a short while, were the most price-competitive on the market, but they are no longer.

I think even Apple knows hardly anyone is in the market for a $7,000 MacBook Pro with a modest amount of memory and storage. From Mark Gurman at Bloomberg, just a while after the news broke:

The company, currently on its M5 series, plans to debut a base M6 processor as early as this year for entry-level Macs, according to people with knowledge of the matter. But in a first, the company will skip higher-end versions of that chip, said the people, who asked not to be identified because the plans are private.

Apple instead aims to introduce its next Pro and Max chips with more advanced computing and graphics power in 2027 as part of a new M7 generation, according to the people.

Apple was, and maybe still is, heavily rumored to announce new high-end MacBooks Pro with organic-LED touchscreens in the fall. But it appears Apple has either gotten into a fight with Taiwan Semiconductor Manufacturing Company over pricing and/or determined it won’t be able to ship enough units at these updated prices for manufacturing them to be worth it. Either Apple fully delays the new MacBooks Pro to late 2027, which would be truly remarkable as it would be over an 18-month gap between revisions, or it will ship them with M5 Pro and M5 Max processors this fall anyway. It’s undeniable that the price increases — both on the supplier and consumer sides — will have some effect on product timelines in the short term.

One final note on this debacle: The RAM problem will likely get worse before it gets better, and I can see this sparking massive backlash to artificial intelligence development. It is hard to describe an aspect of modern life that AI-powered chatbots have not shattered: price increases for computers and video game consoles, a horrible job market, and a loss of creativity in media are all notable consequences of AI companies’ incompetent (and uncharismatic) leadership and inflated valuations. If this isn’t the final nail in the coffin for the AI industry, I’m not really sure what is.

A correction was made on June 25, 2026, at 9:10 p.m.: An earlier version of this article stated that the M5 Max MacBook Pro’s price increased by $600. It was actually $500. I regret the error.

New York Times: Meta Is Building a Gambling ‘Prediction Market’

Eshu Marneedi — Tue, 23 Jun 2026 19:00:10 -0400

Mike Isaac and David Yaffe-Bellany, reporting for The New York Times:

Polymarket and Kalshi, prediction markets where users can bet on outcomes as varied as the Super Bowl and the length of the State of the Union address, have been some of the fastest-growing destinations on the internet.

Mark Zuckerberg has noticed — and he wants in on the action.

Mr. Zuckerberg, the chief executive of Meta, recently dispatched a small team at his company to create a smartphone app similar to Polymarket and Kalshi, two employees with knowledge of the matter said. Users would not wager money, and the app would probably rely on a video-game-like points system instead, one person said, though the company had not ruled out the eventual use of real money betting.

My first thought was that Zuckerberg is the most blatant and prolific intellectual property thief in Silicon Valley. But that’s hardly the most important news here: The world’s most famous social media company is about to launch a gambling app for billions of people.

Prediction markets have taken hold of the United States unlike any modern social product. The FIFA World Cup broadcast is littered with advertisements from Polymarket and Kalshi, and in states where they’re legal, sports betting companies buy out nearly all of the mid-match advertising slots. America clearly has a gambling problem. The premise is somewhat enticing: bet on real events, like sports or politics, and win some money. These “markets” require no experience or specialized knowledge and can potentially pay out thousands of dollars. Kalshi and Polymarket have even positioned themselves as an alternative to political polling by littering the internet with shady, scantly disclosed news sponsorships.

But this is about Meta, one of the richest and most disreputable companies in Silicon Valley. If Meta wanted, it could buy Kalshi and Polymarket, turn them into a billion-user monopoly, and turn the globe into a mess of young gamblers in a matter of months. Knowing Zuckerberg and his history of building snide gimmicks into Instagram and Facebook — so much so that his company lost a lawsuit against those gimmicks — I wouldn’t be surprised if this is the game plan for “Arena,” this new prediction market app. Meta’s apps have always functioned as high-risk, high-reward social platforms, and Arena would be yet another.

The “points system” will probably be how the app works, but that’s not to say real money won’t be involved. (Real money will always be involved; points are useless for Meta’s flailing stock price.) I think Meta would sell those points for a fee through an in-app purchase, or perhaps require people to watch advertisements to earn more. This is how questionable “free-to-play” mobile games became the crown jewel of the App Store — they litter their apps with dark patterns to force people to buy “gems” or “coins.” This is clearly Meta’s overall business strategy, leveraging the wealth of information it has on people from their Instagram and Facebook usage patterns to devise clever bait.

I wouldn’t be surprised if this app launches in the fall, right as the midterm election season begins. It only seems natural for a company like Meta to use its inordinate power to sway an election that could have disastrous consequences for the company.

Tim Cook Says Apple Will Raise Prices Due to Memory Shortage

Eshu Marneedi — Wed, 17 Jun 2026 18:25:15 -0400

Rolfe Winkler, reporting for The Wall Street Journal (Apple News+):

Apple plans to raise prices on its products to offset the surging costs of memory and storage chips, Chief Executive Tim Cook said in an exclusive interview with The Wall Street Journal.

“Unfortunately, price increases are unavoidable,” he said. “We’re doing our best to mitigate the huge increases that are being passed to us, and we’ve been trying to shield our customers from the increases, but the situation has become unsustainable.”

Cook declined to offer details on the timing or scale of the planned price increases, nor which products will be affected. Apple’s next major product launch is likely to be in September when it releases the iPhone 18 lineup, expected to include a new foldable iPhone.

The memory shortage in particular has been debilitating for the tech industry. A new Raspberry Pi — a small computer with 4 gigabytes of memory — costs $150. Similar models used to cost around $100. The shortage, driven by a supply-and-demand issue of RAM manufacturers diverting resources to artificial intelligence servers, has already hit Apple. Earlier this year, it was forced to discontinue the base-model Mac mini and high-end configurations of the Mac Studio with 128 gigabytes of memory and higher storage capacity. It is unprecedented for Apple to flat-out discontinue an existing computer with no mention of its return — Cook’s supply chain prowess ensures that even high-demand configurations will always be in stock at some point. (Just take the MacBook Neo as an example.)

There are two basic, first-year economics solutions to this: decrease demand or increase supply. One is better for consumers than the other. If the AI bubble pops after the initial public offerings of the major AI companies, RAM prices will steadily decrease as data center construction grinds to a halt and the demand for inference slows. Perhaps this could even happen sooner than we initially thought because OpenAI can’t afford to build new data centers. If the bubble never pops, however, we’ll see more capital investment in RAM and storage production to meet demand, and prices may slowly come down. The former situation will have a much more immediate, drastic effect, however, as there’s not much incentive for RAM manufacturers to lower their prices at all if demand keeps rising. (The RAM industry operates as an oligopoly, and increased demand for inference only encourages it to form a cartel.)

While the game plays out, this ought to be quite unfortunate for iPhone and Mac buyers in the fall. I expect iPhone prices to go up by at least $100 in September — and a real price increase at that, not a storage bump like last year due to tariffs. The new MacBook Pro (or, maybe, MacBook Ultra) will probably see even steeper price hikes, though Apple may try to justify them with the new organic-LED touchscreen and new design. And I also wouldn’t be surprised if Apple discontinues more base-model products, like the cheapest MacBook Neo, older-model iPhones, and base-model iMac. If the price of RAM and storage really begins to devour Apple’s margins, I could even see the base-model MacBook Air being discontinued. All of these changes will be unfortunate.

Commerce Dept. Orders Anthropic to Suspend Access to Claude Fable 5

Eshu Marneedi — Tue, 16 Jun 2026 19:39:29 -0400

Hugo Lowell, Lily Hay Newman, and Maxwell Zeff, reporting for Wired (Apple News+):

Trump administration officials concluded talks with Anthropic on Monday without lifting export controls that were imposed last week on the company’s most advanced AI models in response to jailbreaking concerns, according to three people briefed on the matter.

The administration continues to believe that there are ways to disable some of the guardrails on Anthropic’s Claude Fable 5, effectively allowing users to access the more powerful cybersecurity capabilities of the company’s Mythos model, the people said.

Anthropic has said for days that the administration’s concerns are overblown, a position it reiterated in working group meetings held at the Commerce Department with government researchers from the Center for AI Standards and Innovation and the Office of the National Cyber Director, Sean Cairncross, one of the people said.

There are a few angles to this conundrum:

Why did the Commerce Department suspend Claude Fable 5?
Was Anthropic’s “fearmongering” responsible?
Does the Trump administration have the legal right to impose export controls in this way?

First, why? The Trump administration’s official word is that the guardrails on Claude Fable 5 — which prevent users from using the model for certain biological and cybersecurity problems — can be easily jailbroken. Amazon, one of Anthropic’s key inference and investment partners, discovered the guardrails and, instead of reporting them to Anthropic directly, went straight to the Commerce Department. This is uneducated speculation, but that sounds like intentional sabotage to me. Either way, the Trump administration, after hearing about the jailbreak, ordered Anthropic to suspend Claude Fable 5 for all non-U.S. nationals, including those legally in the United States.

This reaction is nonsensical: For one, it is virtually impossible to enforce a domestic export control like this. (I’ll elaborate on this in a moment.) Second, Americans are absolutely capable of sabotaging U.S. military action and compromising national security. What stops a foreign adversary from bribing an American who has access to Claude Fable 5? So the rationale for this is clearly not national security or genuine concern, and fortunately, the Trump administration isn’t tight-lipped on culture-war issues. Defense Secretary Pete Hegseth posted this incredible message on X right after the export controls were hastily announced:

Three months ago, @DeptofWar kicked @AnthropicAI out of our building—forever.

Every passing day proves why that was the right move.

The U.S. government is waging a war against an American company. Media outlets have been caught up in the technical detail of the export controls — and, to an extent, so have Anthropic representatives negotiating — because they have “Silicon Valley brain.” Washington journalists know what’s happening here: the Trump administration is exacting revenge after Anthropic refused to let the Defense Department use Claude models for “all lawful purposes.” There is zero ambiguity here for why the Commerce Department decided to levy such a punitive, impossible punishment. This is not a matter of national security — this is the defense secretary laying it all out in the open.

So I think that makes the second angle moot, but only to an extent: Was Anthropic’s “fearmongering” about Claude Mythos responsible? I think yes, but not in the most direct way. Regardless of whether Anthropic said Claude Mythos and Claude Fable 5 were dangerous or not, they’re the most capable large language models ever created, in every benchmark. Third-party security researchers would inevitably try to jailbreak them because of their capability, not Anthropic’s advertising. But Anthropic has also advocated heavily for government oversight under the Trump administration, with Dario Amodei, Anthropic’s chief executive, saying the administration should have the power to cancel the deployment of its models if it deemed such action necessary. This is the only argument from the pro-Trump administration camp I find convincing: Anthropic wanted this. From Amodei’s essay titled “Policy on the AI Exponential”:

Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety. I am grateful to see the Trump administration’s Executive Order move incrementally towards a greater role for government in AI, though Anthropic’s proposal recommends even further action.

There are really two ways to approach artificial intelligence regulation: the OpenAI approach and the Anthropic one. OpenAI has no morals, ethics, or responsibilities. Its only goal is to make as much money as possible and turn the Great Plains into one gigantic data center. If the U.S. military wants unfettered access to OpenAI models, OpenAI will happily provide it. If the Trump administration wants to tune out “woke” information in the models’ weights, OpenAI will oblige gracefully. OpenAI will do anything to score an extra government contract and inch one step closer to world domination. It has always been that way.

Historically — before this passage in Amodei’s essay — the Anthropic approach has been quite meritocratic. It welcomed the government’s thoughts — particularly around protecting users’ economic and civil rights — but ultimately, it felt the final decisions of how those models were made and distributed should be in the hands of experts, not the government. Anthropic would rather lose government contracts worth billions of dollars than have its models make autonomous decisions in war — that was its red line and it refused to cross it. But after the kerfuffle with the Defense Department in February, I feel Anthropic emerged unnerved. It learned what happens to companies that oppose the administration. It silently abandoned this staunch authoritarian stance and loosened its grip on its models. So now, we’re here, where Anthropic is split between controlling its own models and letting the government regulate them. Anthropic had a good thing going and allowed the U.S. government to destroy it.

So, the third angle: Is any of this even legal? I am not a lawyer, but I err toward no. The government appears to be relying on the Export Administration Regulations, a set of laws that govern export controls. The EAR affects products that have dual military and civilian uses, such as LLMs. The government then imposes an export restriction on a usually narrow scope of countries, like China or North Korea. These restrictions may also be applied to foreign nationals from those countries. The Commerce Department has instead used the EAR to issue a blanket ban on all foreign use of Claude Fable 5, including many of Anthropic’s own employees who worked on the model. This is an outlandish interpretation of this law.

It hasn’t been well-conceived, either. Do dual citizens count? And how does an export restriction in the first place ensure foreign nationals don’t use the model? How does a company enforce such a restriction domestically? The most likely timeline is that the government wanted to punish Anthropic by kneecapping its flagship model, realized it didn’t have a law that allowed it to — thanks to the First Amendment — and used export controls inappropriately to achieve the same effect. I reckon that if Anthropic sued the government, it would win at least a temporary injunction. (Again, I am not a lawyer.) Banning people legally in the United States from using a private company’s technology is truly unprecedented. This whole thing is unprecedented — but it might just be Anthropic’s fault.

Siri AI, Along With the Rest of Apple Intelligence, Is Initially Unavailable In EU

Eshu Marneedi — Fri, 12 Jun 2026 00:41:23 -0400

Apple Newsroom, shortly after this year’s Worldwide Developers Conference:

Apple today introduced Siri AI, an entirely new version of Siri, powered by Apple Intelligence. Unfortunately, due to the Digital Markets Act (DMA), Apple will not be able to ship Siri AI in the European Union with the release of iOS 27 and iPadOS 27. Over the past several months, EU regulators did not accept any of Apple’s proposed solutions to bring Siri AI to the EU while safely supporting other virtual assistants.

“We’re deeply disappointed that our EU users won’t have Siri AI on iPhone or iPad when we share our new software releases later this year,” said Craig Federighi, Apple’s senior vice president of Software Engineering. “Our hope is to eventually bring Siri AI to the EU, and we will continue to engage with EU regulators on a path forward. However, their refusal to engage constructively on solutions that preserve privacy and security means we do not currently have a timeline for Siri AI’s availability on iOS and iPadOS in the EU.”

The DMA — or, at least, the European Commission’s interpretation of it — demands Apple allow any external large language model-powered chatbot access to the same information and plugins that Siri AI does. Siri AI has, among many other tools, access to the app toolbox and semantic index to use a person’s personal context and App Intents to personalize responses. Third-party LLM extensions, such as ChatGPT’s existing one, do not have access to these features, as there is no practical way to keep that data on-device if it were offered to third parties.

And yes, Siri AI does support third-party extensions, though they’re (probably intentionally) quite barebones. People can really only just chat to external chatbots through the Siri interface — they can’t do most of the things Siri AI can. The only reason this is even in the operating systems is most likely to stave off regulatory pressure from both the commission and the U.S. government, which filed an antitrust lawsuit against Apple during the Biden administration in 2024. But clearly, for the European Union, it isn’t enough. So why isn’t it?

The short answer is that I think the European Union doesn’t know what Apple is even talking about. They’re too decrepit to understand the intention of their law, how Apple is trying to comply with it, and what’s best for consumers. They, frankly, just want Apple to “figure it out.” The commission doesn’t fundamentally understand what on-device inference is, how Private Cloud Compute keeps user data private, and how Apple and OpenAI’s approaches to artificial intelligence differ. I don’t believe they think it’s all the same, but more like they think it’s all just kids playing in the yard. They can’t possibly be bothered with this “nerd stuff.”

Apple does not have access to people’s personal data. It is end-to-end encrypted if they so choose — a feature both the European Union and the United Kingdom have tried to eliminate — and always off-limits if a person decides not to store it in iCloud. Apple could not, even if it wanted to, read someone’s on-device Apple Notes. Siri AI can read those notes because they usually don’t need to be sent to Apple’s servers for processing — the inference is also done on-device. And if the inference does need to be sent to the cloud, it is encrypted in transit, decrypted using a special two-way handshake between the client and the server, and immediately erased the second it leaves the server. Apple could not, even if it wanted to, peek into the server.

By contrast, a request to OpenAI’s servers is very different. None of it is encrypted or on-device. As soon as a person hits Return on a query, it is sent to OpenAI’s servers, stored there for at least 30 days — if not longer — and trained on by default to improve future models. Human reviewers at OpenAI regularly monitor conversations and suspend users’ accounts for terms-of-service violations. Nothing about the ChatGPT experience is private — OpenAI itself tells its users not to share sensitive data with ChatGPT. If Apple gave ChatGPT access to a user’s private, on-device notes, it would be perhaps the biggest privacy scandal of the modern era.

Unlike the original suite of Apple Intelligence features, I truly think Europeans are missing out on something incredible here. Siri AI is the most private, environmentally conscious AI system on the planet. These are not considerations for a place like China — where Apple Intelligence is also unavailable due to regulation — but they seem like they would strongly appeal to Europe’s abundantly liberal politics. I don’t expect protests on the street from angry Belgians or whatever, but Siri AI seems like the first AI system really made for Europeans. I strongly respect Europe’s commitment to regulation, and in some ways, I think Apple does, too. That’s why the European Union’s regulatory action got a whole press release, while China’s only received a footnote. Europe is a liberal democracy worth fighting for.

macOS 27 Golden Gate Reverses Awful Menu Item Icons

Eshu Marneedi — Thu, 11 Jun 2026 15:50:44 -0400

Tim Hardwick, reporting for MacRumors:

In macOS 27 Golden Gate, Apple has removed many of the menu item icons that are so prevalent throughout macOS 26 Tahoe, as spotted by Nikita “Tonsky” Prokopov (via Daring Fireball). The developer shared before-and-after screenshots on Mastodon to evidence the reversal.

Tahoe was the first version of macOS to place a small icon next to nearly every entry in the menu bar across Apple’s apps, but the change drew swift criticism from designers and developers. Many of the icons are inconsistent and often difficult to understand on their own, with different Apple apps showing different icons for the same menu items.

Most system menus in macOS Golden Gate no longer have icons at all, and the ones that do are consistent across system apps. For instance, the Share button always has a share icon (square.and.arrow.up). Third-party apps that previously adopted menu item icons no longer display them on macOS Golden Gate — developers must programmatically specify that icons are required for particular menu items. (They still appear in macOS 26 Tahoe, even when compiled on the macOS Golden Gate software development kit.) Apple appears to adopt icons for major controls: Passwords and Finder use them for views listed in the sidebar, for instance.

I suppose most Mac developers will leave all icons off because that’s now the default behavior. (To override this behavior, as Apple does, developers should use .labelStyle(.titleAndIcon) in SwiftUI and NSMenuItem.preferredImageVisibility in AppKit.) Developers who do choose to adopt the new guidance will have to do it with intention. Every app will be different and choose to use icons in different menu items. This is not an inconsistency — it appreciates that every app is different and requires various user interface considerations. The old icons were ambiguous because they assumed every app needed an icon for every menu. The new system permits choice, and, ergo, tasteful UI design. And it’s not like apps that choose to eschew icons entirely will look dated, either.

The same goes for the new, tighter corner radius. All apps, regardless of their UI framework, use a single corner radius. It looks amazing in macOS Golden Gate — one of those tweaks you recognize as soon as you boot up the computer for the first time. macOS Tahoe looked like it was designed by a child with Fisher-Price blobs and rounded rectangles throughout. The new windows, combined with the more Aqua-like Liquid Glass elements, feel “grown-up.” They look like they’ve been designed with care and by someone who knows something — anything — about UI design. Again, it’s instantly recognizable and striking.

There are lots of other tweaks, too, like the new sidebars that no longer float over the window. They’re just great. They look every bit as Liquid Glass as the previous version, but they feel like they take up less space. Your eyes are forced to focus on fewer specular highlights and disoriented corner radii. Fundamentally, there’s less happening in the interface. Toolbars just look normal again — they don’t float over the content as they did in macOS Tahoe. The toolbar is unabashedly a toolbar, and the sidebar is unashamedly a sidebar. Whoever thought elements explicitly marked as “bars” should be part of the main content was not thinking straight. macOS Golden Gate might be my favorite operating system this year.

At WWDC 2026, Siri AI Aims to Make Amends

Eshu Marneedi — Tue, 09 Jun 2026 20:59:11 -0400

The ‘more personalized Siri’ is finally here — and it’s worth its salt

Siri AI is here. Image: Apple.

Hey Siri, play “We Are the Champions” by Queen.

On Monday, live from its Worldwide Developers Conference in Cupertino, California, Apple’s reorganized artificial intelligence division had one noble goal: Make Siri a champion. The new, completely revamped Siri, dubbed Siri AI, has been two years in the making since its initial announcement at WWDC 2024. Since then, the colloquially named “more personalized Siri” has suffered setback after setback, initially slated for release in spring 2025 before being postponed to the following spring, and finally to the next generation of Apple’s operating systems. Much of the old Siri team is out — Robby Walker, the leader of the Siri division, resigned; John Giannandrea, Apple’s machine learning chief, whom the company poached from Google, was unceremoniously stripped of his duties and retired.

Siri AI realizes the vision Apple had for Apple Intelligence all along: a suite of frontier-class large language models, running mainly on-device, that can handle web searches, tool calling, and agentic work in a user interface familiar to ChatGPT users. Siri AI might be the flagship product, but Apple executives are eager to emphasize that Apple Intelligence is flexible and extensible for the company’s future AI endeavors. Siri AI is fast, powerful, private, and perfectly Apple-like. It turns a product genre almost universally besmirched for its “sloppiness” into a service worth using. And none of it is vaporware — it all shipped Monday for app developers to integrate with.

Expectations are sky-high for Siri AI and the slew of other Apple Intelligence announcements unveiled during Monday’s keynote: everything from the new system-wide dictation models that instantly vaporized a class of Silicon Valley start-ups overnight to the new image manipulation tools that rival Google’s. There is much ground to cover, but in many ways, WWDC 2026 was a return to the status quo for a company that has faced its biggest turmoil in years. After years of anemic developer relations, executive turnover, and a controversial redesign of its operating systems, I’m happy to say that Apple clocked a win this week — and so will the millions of Apple users come September. The Silicon Valley enterprise best known for making great products has, again, shipped a great product. The question now is whether its tardiness is excusable.

The Architecture

There was a not-so-distant moment when we all thought none of this would ever ship — when the story of the new “more personalized Siri” seemed destined to resemble that of the AirPower — unveiled at a high-profile keynote, delayed, and never to be spoken of again. It didn’t help that, concurrently, a deluge of rumors was swirling about Apple’s AI division and leadership. Employees were directionless, progress seemed stagnant, and new bugs seemed to stymie the project’s launch for years. At the core of the new Siri project, unveiled at WWDC two years ago, was an LLM that called App Intents actions in apps, leveraging a so-called “personal context” gathered through a semantic indexing feature. Each of these distinct parts suffered setbacks.

For one, the LLMs of the time — let alone the comparatively limited ones Apple had developed amid a lack of funds — were not advanced enough to reliably call tools. To meet expectations, the model must be post-trained on those tools and usually undergoes a process called reinforcement learning to reward the model when it correctly calls a tool. These tools — like web search and code interpretation — are exposed as instructions in a hidden section at the beginning of a chat. Apple wanted to expose App Intents, a protocol that allows apps to donate actions to the system, as tools for the model, but Apple’s models at the time were insufficiently trained to call these tools. The resulting product was practically unusable because it never reliably performed actions within apps, a marquee feature of the new Siri.

The new Siri also relied heavily on personal information to distinguish itself. While OpenAI and Anthropic were forced to rely on public application programming interfaces, or links to system data, Apple and Google could leverage personal data they already knew about users. Apple’s solution was to rely on a new framework called the personal context, a collection of user data that the model could tap into at any time. It was an early implementation of memory — a feature now built into every major chatbot — but perhaps more advanced, as it required apps to donate snippets of information. Apple planned to allow apps to contribute information like messages, notes, and emails to the system via the semantic index. But again, the models were too underpowered to know what information they needed for a query — they didn’t know what to look up in the personal context to answer a question.

Two years later, Apple is partnering with Google to distill Gemini models for use in Apple Intelligence.¹ The models work in a new architecture where a classifier model, known as the system orchestrator, determines the type of query and which information and models are needed. The chosen model then calls the tools and returns an answer, usually on-device and occasionally in the cloud. There are three types of information the system orchestrator may deem necessary: the app toolbox, the semantic index, and on-screen context. Once the query is assigned to an adequately sized model, it may choose to use Apple World Knowledge, a pseudo-web search engine developed by Apple for its foundation models.

The app toolbox includes App Intents made by third-party developers. Any action in an app can be an App Intent — they run by silently launching the app in the background. Examples include playing a song, sending a message, or marking a task complete in a to-do list app. App Intents already exist across the system, such as in interactive widgets, Spotlight, and most importantly, Shortcuts, where their breadth is fully appreciated. The semantic index includes data from these apps — corresponding examples include the music in a person’s library, messages people have sent, and tasks in a to-do list. Developers can contribute actions and data through the App Intents framework: actions are called intent schemas, and data points are entity schemas. Apple’s foundation models can also, when requested, capture a screenshot of an app and analyze visual content.

Apple Intelligence uses two on-device models and three cloud models. All five are built “in collaboration with Google” and are pre-trained on Google’s silicon, known as tensor processing units. The two on-device models, AFM Core and AFM Core Advanced, have three and 20 billion parameters, respectively.² The latter is only available on newer Apple devices and is not required to use most Apple Intelligence features — it actively uses only three to four billion parameters at a time through a mixture-of-experts architecture, i.e., the model chooses which parameters are relevant for a query as opposed to loading them all concurrently.³ AFM Core Advanced is also natively multimodal for voice and image processing. Craig Federighi, Apple’s software chief, said most queries rely on these on-device models.

Two of the cloud models are similarly distilled and made in-house by Apple: AFM Cloud and ADM Cloud, an image diffusion model. (ADM Cloud is used for image generation features in Image Playground.) The on-device classifier chooses to send longer, more intricate queries to these models, hosted by Apple’s own Private Cloud Compute servers. These models are specially designed to run on Apple silicon, much like the on-device AFM models. Their size, however, makes them impractical to run on mobile devices. For the most demanding queries, Apple Intelligence uses AFM Cloud Pro, a Google-made model post-trained by Apple that runs in Google Cloud’s Nvidia servers. Apple is, notably, not relying on Google and Nvidia’s privacy stack — these servers will still use Apple’s Private Cloud Compute software, and only Apple is authorized to run inference on them. Apple user data, as Federighi stressed, never touches non-Apple software.

This architecture puts to rest the incessant speculation that Apple would somehow abandon Private Cloud Compute for Google Cloud, which I have always dismissed as a preposterous rumor. Apple Intelligence continues to run on Private Cloud Compute — usually on Apple’s bespoke hardware running Apple’s own models — and most of the time, it relies on on-device processing anyway. (I can confirm this while testing Siri AI.) I believe this nets Apple two crucial wins: privacy and environmentalism. Private Cloud Compute, for the most part, runs on renewable energy, and on-device models are the most energy-efficient way to run inference. And Apple can wax poetic all it wants about how users’ data is never used for training or even seen by the company.

The new models and APIs are designed to work reliably and quickly. Siri AI should be able to understand a query through the classifier, gather the requisite data, send it to an appropriate model either on-device or in the cloud, optionally search the web, and return a result. App developers should be able to donate actions and information to the system. And the models themselves — since they’re distilled from Gemini — should be powerful enough to call tools efficiently and produce accurate responses. The new Apple Intelligence architecture — as complicated as it may seem — appears sufficient to meet those demands.

Siri AI, Née the More Personalized Siri

As Mike Rockwell, the executive under Federighi responsible for the new Siri project, said onstage, Siri AI’s paramount priority is better capability. I’ve written many times that any platform-native AI product must master three modalities: device use, search, and agents. This might seem simple enough, but excelling at all three has proven to be a difficult task for LLMs. The new Alexa+, for example, was fine at search and agents, but regressed in the most basic tasks, like controlling smart home accessories, because LLMs are inherently nondeterministic. The classifier, therefore, must choose which architecture to use: the deterministic one or the lossy one. Or, occasionally, some combination of the two.

Siri AI supports all three modalities. App Intents function as programmatic, lossless gateways to agentic work in apps; the Apple World Knowledge search engine gives models access to reliable, recent information; and Spotlight — now combined with Siri AI — is a reliable interface for file search. When Siri AI doesn’t need an LLM, it doesn’t use one. It uses App Intents, for example, to change settings, search for files, or play music. The model is in front of the interaction, not behind it — it only serves to process and analyze the query before calling a faultless, programmatic App Intent. When Siri does need an LLM, such as to summarize web results, it hands the query off to a more powerful model, perhaps in the cloud, to come back with accurate results.

Users can invoke Siri AI in two main ways: by pressing and holding the Side Button to open voice mode, or by swiping down from the middle of the screen to open the combined Spotlight and Siri AI view from the Dynamic Island. (It is also possible to swipe down on the Dynamic Island, but I have a hunch Apple will remove that after complaints.) On macOS, Siri AI can be invoked through the “Hey Siri” wake word, a system-wide context menu, or Spotlight. Siri effectively resides in Spotlight, and when it detects a query that matches no search results, it kicks off a Siri AI prompt. Spotlight is not replaced by Siri AI; the two features live in congruity. On all platforms, Siri AI has a gorgeous animation when it’s listening to user input or reasoning. It shows a spinning icon of six circles and occasionally makes soft beeping sounds to indicate it is reasoning after a few seconds. Results show in a semi-translucent black-and-clear sheet that takes full advantage of Liquid Glass.

When using voice mode, Siri writes shorter answers, usually one or two sentences. It doesn’t show web search sources visually in the first beta, but it does read them aloud, a minor gripe of mine. The system, as a result of its brevity, is fast — much faster than the keynote demonstration, a classic case of underpromising and overdelivering. Cloud and web search queries take slightly longer to process, but they’re still almost instantaneous compared to chatbots like ChatGPT and Gemini. And unlike Google’s flawed AI Overviews, I have yet to experience a single hallucination or uncanny response. I partially attribute this to the models having little personality: they’re quite dry compared to even Gemini. Users can pull down on the initial answer to view a chat interface and continue the conversation.

Conversations are saved in the initially barebones Siri app. The app truly looks as if it were cobbled together by Claude Code on the eve of WWDC; I assume it’ll improve through the betas. The app shows a monochrome list of conversations, and tapping on one opens a chatbot-style interface with options to attach files or enter voice mode. While the interface looks eerily similar to ChatGPT, it still has all of the same Siri features. People can ask it to, for example, play music or toggle a setting; it isn’t limited to typical chatbot-style conversations. It’s certainly possible, though uncanny, to open a new chat in the Siri app and ask it to turn on dark mode from there — it will happily display the toggle and everything, just like normal Siri.

All of the initially marketed features — including those from 2024 — function in the beta. Siri AI can call existing App Intents from developers who’ve already added them, and it can search the personal context. (The personal context is currently limited to first-party apps, as no developers have added support for it yet.) The system itself is usually quite accurate and triggers the correct tools, but it is sometimes trivial to confuse. I asked the model, “When was my last flight with United Airlines?” and instead of checking my email inbox for a confirmation, it ran the United Airlines app’s App Intent to list recent flights. I had to tell the model that it had been a few months since the flight for it to look in the Mail app. Perhaps a human assistant who didn’t know off the top of their head when my last flight was would have made the same mistake. (I can see the “When does my mom’s flight land?” scenario from 2024 working perfectly in this beta.)

Siri chooses which tools to call and expose in the interface. When an App Intent returns a result, it displays it in-line, just like the previous version of Siri. When an answer only requires text, it prefers headers and paragraphs to bullet points. The model doesn’t expose its reasoning traces or say how long it has reasoned for, and the interface isn’t cluttered at all, remaining focused on the answer and App Intents. It also doesn’t indicate whether inference was done on-device or in the cloud, but all of mine were clearly done on-device — my iPhone 17 Pro warmed up considerably while testing Siri AI. (iOS 27 is otherwise remarkably stable.) If Siri pulls from the personal context, it displays the information — such as an email, text message, or note — in an App Intent-like bubble, which can be tapped to view the source content. (Sources are always displayed at the bottom of the response in the Siri app.)

Web search results and world knowledge are fast and reliable. This is where the new Siri truly feels like a ChatGPT-style product and less like the original version. It’s a full-blown LLM, after all, and it knows everything, often without even having to reach for the web. Chances are, most people accustomed to the free Gemini and ChatGPT models won’t even notice a difference in accuracy when using Siri. (The models are definitely less sophisticated, though.) Siri even passes the “car wash test” when asked, “If a car wash is 5 meters away, should I walk or drive there?” In practice, the model was able to do time zone conversion, simple explanations, and more complex math through its own parameters. The model never lists search results; it always chooses to summarize them in a paragraph.

Apple’s keynote demonstrations, notably, were not aimed at ChatGPT, and I think this is emblematic of the company’s strategy: Siri AI isn’t meant as a replacement or competitor to ChatGPT. OpenAI has the model advantage, but Apple has the ecosystem advantage. There’s a reason that both Gemini and ChatGPT are equally popular — it’s not because there are two billion unique users of AI chatbots altogether, but that the billion people who use ChatGPT are also likely to use Gemini. Those people will probably also use Siri AI — if they own a new-enough iPhone, at least — to do things they’ve always asked Siri for, just with more precision and depth. Siri, for most people, has always been — and fallen short of being — an assistant to handle small tasks on their iPhone, like sending messages or finding emails, and now that experience will be augmented by generative artificial intelligence.

This is the central, axiomatic thesis behind each of Apple’s demonstrations: The appeal of Siri AI is not the models themselves, but what the models enable. The models are simply how Apple gets to a Siri that lives up to people’s needs and desires, and what ultimately suspended the project for two years. By contrast, the appeal of ChatGPT is the models, not the product. There are hundreds of products just like ChatGPT on the App Store, but what keeps OpenAI at the top of the list, even among its most direct competitors, is its models. The models are OpenAI’s moat; the data is Apple’s moat. Apple, unlike OpenAI, doesn’t have to attract new developers to build Model Context Protocol-powered apps. It doesn’t need them to build GPTs or skills or plugins or whatever other gimmick OpenAI cooks up to charm users. It has the App Store, a moat so brilliant that it has been sued countless times on nearly every continent for its brilliance.

People won’t go to Siri AI for writing advice or therapy. It won’t be the most powerful research tool, write software, or solve the next Erdős problem. The distilled Gemini models aren’t there for the sake of being great models — they’re there to leverage Apple’s idiosyncratic moat. I don’t think the ramifications of Siri AI will be felt in Silicon Valley writ large because it will not replace the usage of ChatGPT, Gemini, or Claude. The models just aren’t there, and they don’t have to be for Siri AI to succeed. I can see some Google Search market share fading due to Siri AI’s placement in Spotlight, but (a) as I wrote last month, Google wants that to happen anyway, and (b) the volume of Google searches is so high that it’ll barely make a dent. Siri AI is for the people who use their Apple products all the time and just want Siri to work.

Concurrently, I think Silicon Valley misconstrues just how large that contingent is. Siri, in its current, nearly-inept state, processes billions of server-side requests. People ask Siri legitimately difficult questions and hope for useful answers. If I had to guess, most of those questions are web search-related, and the new Siri is fantastic at synthesizing content from the web. All LLMs are, and the new Siri does it quickly and on-device. The Valley has spent Monday underselling the new Siri’s utility because it lives in its own hive mind, detached from the rest of humanity. Nobody cares about coding agents and “loops,” whatever they are — they just want to know when the next New York Knicks game is and when Subway closes. And believe me, the volume of people asking ChatGPT about the Knicks game on Monday is tenfold the number of people who have ever heard of Codex. Apple is not behind on anything meaningful; the rest of the Valley is putting the cart before the horse.

This is an unwonted claim from me, but I think Siri AI does almost everything it must. It’s amazing at quick searches, App Intents are smart and more useful than anything OpenAI has ever unenthusiastically shoved into ChatGPT, and it’s a good voice assistant. There’s very little slop here, and, dare I say, the vibes are quite phenomenal. Apple has always been a product company first, not a research lab, and Siri AI embodies that ethos strongly. Apple doesn’t even own the core technology — the Gemini models it distilled to make these Apple Intelligence features possible — but it has made a product that is somehow more compelling than much of even Google I/O this year. Apple is indubitably “behind” on AI because there are truly only two companies ahead: Anthropic and OpenAI, but I’d argue that catching up to them is a trapdoor anyway.

As I wrote in 2024, the developer story still has a metaphorical question mark looming over it, but I think major developers are more inclined to participate in AI features now than two years ago. The industry has diversified considerably since then, with Google and Anthropic now owning significant shares of the market. It would be petulant for major corporations like Uber, Google, and Microsoft to offer extensions for ChatGPT and Claude but snub Apple, whose user base theoretically extends far beyond any number OpenAI could ever hope to reach. I attributed my uneasiness in 2024 to developer tensions after the lackluster launch of Apple Vision Pro and the Digital Markets Act in the European Union, but despite stagnation in that part of Apple’s business, I think developers will be forced to act this time. There will be people who want Siri AI to work in all of their apps.

If the developer story has a happy ending, Apple would, in many ways, take the lead in the AI war. Start-ups and “Big Tech” have tried to teach vision models how to use computers natively and have gotten close in recent months, but those models are still nascent. App Intents are a programmatic alternative that don’t require an ultra-powerful model but can provide an equal or better user experience. (Anthropic and OpenAI also concluded that programmatic interfaces, like MCP servers and command-line interfaces, are more reliable than computer use, but with little broad support.) People could, all of a sudden, do legitimate agentic work on their device using the new Siri, like place orders, organize files, or send emails. I know from experience that App Intents are tricky for developers but not onerous to get right. I say every developer should hand the task off to an agent and see what happens.

The overarching theme behind WWDC this year is that this is real software for real people. It’s irrefutable that this is the kind of breakthrough that the AI industry, after years of insular thinking, needs to resonate with normal people asking the Valley what benefit LLMs and data centers have to them. No, the new Siri isn’t running on the most powerful models, but it’s a vastly more eloquent solution than anything Google has ever shipped, and it does the job prescribed two years ago dutifully, in Beta 1. There’s a long road ahead for Siri AI, but Apple is now where it needs to be, as long as third-party developer adoption remains strong. I’m pleased enough to give Apple a pass for now, not because it’s leaps and bounds ahead — or even because its tardiness is excusable — but because it has potential. Really strong potential.

The Rest

Siri AI and Apple Intelligence aren’t the same: Siri AI is a feature of Apple Intelligence. The models are integrated throughout the system in more helpful and creative ways, much more so than the initial 2024 set of releases. Existing features have been improved with the new models, such as Visual Intelligence and Writing Tools, both of which are now powered by Siri. In most text fields across the system — on iOS, iPadOS, macOS, and even visionOS — Writing Tools has been rebranded as Write with Siri, using the new models to draft emails, text messages, and notes, sometimes in a person’s own writing style. The Proofread feature uses a new user interface to underline and review changes, much like Grammarly. These are table-stakes improvements to features that languished during the first Apple Intelligence rollout, and I’m interested to try them throughout the betas.

Visual Intelligence is also baked into the Camera app, labeled as Siri mode. People can take photos of meals to get an estimate of their nutritional value and easily log them in Health; split bills by snapping a photo of the check; or import passes into the Wallet app. I assume the feature set will grow over time, similar to Google Lens. (Visual Intelligence still has the option to search Google, along with other third-party apps.) These features are also available when taking a screenshot or viewing a photo later. I think of Visual Intelligence as Apple’s consumer-facing name for image analysis, whereas Siri is the interface that displays answers. You open Siri mode, Visual Intelligence analyzes the image, and Siri gives you a response. Siri itself uses Visual Intelligence in the background to analyze photo uploads and for the on-screen awareness features. It even works on Apple Vision Pro — just look at something, either on-screen or in person, and ask Siri about it. I wonder what that could be for.

One of the most interesting underlying stories at WWDC this year was third-party model adoption. Apple doesn’t just want third-party developers to stop at supporting App Schemas — it wants them to use Apple’s foundation models in their own apps, either on-device or through Private Cloud Compute. Apple laid the groundwork for running its models in third-party apps last year, but this year, the models are far more powerful, especially now that they can run in the cloud. And if developers want to use third-party models, like those from Google or Anthropic, the company has updated the Foundation Models framework to simplify API calls to those providers, in addition to Apple’s models. (OpenAI was interestingly not mentioned, but it can add support for the framework if it wants.) If developers want to run third-party, open-source models on-device, they can use the new CoreAI framework to run efficient inference on Apple silicon.

Apple has integrated its own foundation models throughout various apps. In Safari, Apple Intelligence can group tabs by topic and monitor webpages server-side for specified updates. Users can also prompt the models to write bespoke Safari extensions. These are the kinds of updates that needed powerful, cloud-based models, and now the foundation models can write code with competence. They won’t be as fast or powerful as Claude Code, and the code certainly won’t be shippable to the App Store, but it’s amusing to see Apple embrace the personal software revolution this way. People can also prompt the system to create shortcuts in the Shortcuts app on their behalf. Powerful models make a difference.

Perhaps the most pertinent example of the models is the new agentic Automatically Fix Passwords feature. Current versions of iOS alert users to compromised or reused passwords in the Passwords app, but changing them has always been a bit of a hassle. On Apple Intelligence devices running iOS 27, Passwords uses an on-device computer-use model to log into the account, navigate to the Change Password menu, change the password using the Passwords app’s suggestion, and save the new one. It even collects two-factor codes from Messages or Mail if needed. I think this is such an ingenious application of computer use models, not only because the feature is plainly fascinating, but because it will improve security for millions of Apple users. I think every third-party developer ought to play around with the new models, for the sake of their users.

But playing catch-up has its consequences. In 2024, about Image Playground, I wrote:

I’m less concerned about the social justice angle many have seemed to stake their beliefs in and more about the feelings this feature creates. Apple users, engineers, and designers all share the conviction that software should be beautiful, elegant, and inspiring, but oftentimes, the wishes of shareholders eclipse that unwaveringly essential ideal. This is one such occurrence of that eclipse — a misstep in the eyes of engineers and designers, but a benison to the pockets of investors. Apple has calculated the potential uproar within a relatively and probably measurably minor slice of its user base isn’t worth it in favor of the deep monetary incentives, and it worked for the C-suite executives. Will Image Playground and Genmoji change the way people use and feel about their devices? Possibly, maybe for the best, or maybe for the worse — but what it will do with resolute certainty is upend the value of digital artwork.

Image Playground was contrived in the most un-Apple-like manner. It was never a good feature to begin with; hardly anyone used it after it was updated to use OpenAI’s image generation model, and AI-generated art is still overwhelmingly loathed. The foundational idea that someone can generate AI images of another person in their contacts is so beneath the Apple ethos. People legitimately and rightfully hold Apple to a higher standard than Google and certainly OpenAI. In an ideal world, Apple would’ve scuttled the feature after internal backlash, but alas, it has chosen to unscrupulously double down.

Image Playground in iOS 27 returns to using Apple’s own models, this time with new styles, including — try not to gag — “photorealistic.” I cannot believe Apple’s clearly tasteless executives have stooped so low that they now find it acceptable to put a photorealistic AI image generator that can create photos of real people in the pockets of hundreds of millions of users. Ideals be defenestrated — how has Apple not considered the backlash to AI image generators? It is not 2024, and AI image generators have become mainstream. The contingent advocating for their removal is not a minority, but clearly within Apple it somehow is. For a company once hellbent on producing products at the intersection of technology and liberal arts, it seems like it has eschewed its driving principle. Image Playground is unbefitting of Apple.

Perhaps that is not even the most objectionable of the image generation features announced this year. The Tools tab of the Edit menu in Photos now has three options: Clean Up, Extend, and Reframe. Extend and Reframe are the two additions, stolen straight from Google’s playbook: Extend generates new content to enlarge a photo after it has been taken, and Reframe allows people to adjust the angle and position of the camera, filling in gaps. Both of these tools muddy the waters and blur the line between photography and computer-generated imagery. A photo modified in this way after the fact does not exist — it is definitionally not a photo. Neither is it art, as it is created by a computer, not a human being with the skill and soul to produce a new work. The gall to produce something like this and call it a photo.

To a notable extent, I find Clean Up repugnant in much the same way, but at least in that case, the intent is to remove, not add. Additive content is intentionally misleading because it argues against the core principle of photography: to capture a moment in time. What good is that capture if it can be manipulated in whatever way by someone with no artistic ability? What is the utility of that capture if it is modified by a computer with no intent? This is not Photoshop; it encourages deception. There is no barrier to entry. As this technology develops, it inherently becomes more deceptive. This is a death spiral that nearly every company and government in the world is trying to combat through laws and rules — companies have been sued for it — and for some inscrutable reason, Apple chose now as a great time to go headfirst into the arena.

I don’t intend to belabor this point, but the litany of societal and philosophical issues around the development of generative AI is perennial. When Federighi, the software chief, introduced the Siri portion of Monday’s keynote, he plainly stated: “Still, some appear to be racing forward, seemingly pursuing AI for the sake of AI without clear regard for the people — all of us — that it’s ultimately meant to serve. At Apple, our mission has always been to turn the potential of advanced technology into helpful and intuitive products for everyone.” And what an evergreen statement that is. That philosophy has ostensibly guided all of Apple’s AI features announced on Monday except the image generation ones, and that’s a real pity. I’d love for one Apple executive to compellingly explain how they think image generation is positively impactful to society — I think they’d fall short.

Time will tell how well Monday ages in the history of notable days, but I think it has certainly earned its place on the list. There is much to do, both for Apple and developers, but Siri AI is a promising, hopeful, though admittedly overdue start. I think that’s a good place to be — in a way, that’s where the rest of Silicon Valley is. It’s ignorant to disregard the genuine uses of generative AI in 2026, and Apple fortunately hasn’t, but it’s simultaneously important to be astutely aware of its complications. Image Playground and the three photo editing tools debase Apple’s core philosophy, while Siri AI and useful features like the agentic Passwords tool clearly embody it. It’s about to be an eventful summer of beta releases and third-party apps.

At the beginning of this article, I posed a deceptively straightforward question: Is Apple’s tardiness excusable? Has the company done everything it promised two years ago, plus enough to make up for lost time? It’s probably — nay, definitely — too early to tell, but I’m erring on the side of yes. I think it helps that the initial idea was nothing short of brilliant, and that Apple didn’t get caught in the “hype cycle” of agentic AI, especially last winter. The new team, led by Rockwell, was disciplined, and that’s quite an admirable feat for the Valley. The models and interface are here, and now what’s left is developer adoption and rapid iteration. It’s not like now is the time to kick back and watch the sunset.

Throughout the rest of the year, it’ll also be important to keep an eye on how chatbot usage changes as people begin using Siri AI. Both Anthropic and OpenAI have filed for initial public offerings, poised to be completed later this year. While I think it’s unlikely that Siri AI will eclipse chatbot usage — especially among younger users who use them for schoolwork and programmers who use them professionally — it’ll be interesting to see if people think of Siri as a useful tool worth prompting now and then. I especially see Siri as becoming the de facto web search tool due to its prominent placement in Spotlight. Hallucinations will also be interesting to observe — I’d love to see some frontier model evaluations using the new Foundation Models command-line interface.

Ultimately, I’m relieved that the two-year-long saga has come to a close. The more personalized Siri is, as long as you’re not European or Chinese, finally here. But perhaps that’s just the start of yet another rigmarole.

Distillation is a process where a model is trained on an existing model’s outputs and thinking traces. The child model learns from the patterns of the parent model, exhibiting many of the same behaviors and producing better responses as a result. Most Apple Foundation Models are distilled and made — pre-trained and post-trained — by Apple, not Google. ↩︎
Parameters are a measure of knowledge for machine learning models. ↩︎
There is rampant misinformation about this online. Every device that Apple promised Apple Intelligence would work on — from iPhone 15 Pro onward — will support nearly every new Apple Intelligence feature launched on Monday. The two exceptions are the new voice-to-text dictation model and a new, more expressive voice for Siri. (Users in the European Union and China cannot use Apple Intelligence, much to Apple’s chagrin, irrespective of hardware.) ↩︎

Microsoft and OpenAI

Eshu Marneedi — Thu, 04 Jun 2026 18:15:58 -0400

Hayden Field and Tom Warren, reporting for The Verge:

For years, Microsoft’s AI business leaned hard on its early and exclusive partnership with OpenAI. But the drama-filled marriage slowly devolved into a situationship, and the pair effectively separated in late April (though Microsoft is still OpenAI’s primary cloud partner — for now). This year’s Build had the vibe of a freshly single divorcée posting a thirst trap on Instagram. “It’s always fun to be at developer conferences in times of great change,” Microsoft CEO Satya Nadella said onstage Tuesday, adding that events like this are about “coming to grips with the new opportunity.”

AI chief Mustafa Suleyman, in an interview with The Verge, put it even more bluntly.

“The goal is to prove that we can become one of the top four labs in the world,” Suleyman said. “There’s three labs that matter, Google DeepMind, OpenAI, and Anthropic. We are not one of them at the moment, and that’s always been my intention. It’s why I came here. I want to build the very best frontier models in the world, fully multimodal, and in order to do that, we have to prove that we can do everything that we need to from the ground up, and we’re not just going to take from others.”

One of Microsoft’s first steps at Build was indeed to play catch-up on AI models. Suleyman unveiled MAI-Thinking-1, the company’s first reasoning model, along with six other new models focused on image, voice, transcription, and coding. Microsoft said the medium-size MAI-Thinking-1 model, which will likely be marketed to primarily enterprise clients, is “built from scratch for serious math, coding, and real-world enterprise deployment.” Microsoft is years behind both OpenAI and Anthropic here; OpenAI began releasing reasoning models in the fall of 2024. But Suleyman emphasized its performance on benchmarks like coding and its price point, saying it was cheaper than OpenAI equivalents on some tasks — a big deal in the age of the AI money squeeze, which has inspired a lot of complaints with customers.

I think the main takeaway from this story is that Microsoft desperately wants to be known for anything other than its consumer products, Windows and Microsoft 365 (née Microsoft Office). That’s why it invested in OpenAI through Azure usage credits. Satya Nadella, Microsoft’s chief executive, has concentrated his company’s efforts on cloud computing, and when ChatGPT launched, Microsoft had no real choice but to back OpenAI and become the inference provider. As for OpenAI’s end of the deal, Microsoft integrated GPT-4 into Bing and Windows under the Copilot suite of products. (Remember Sydney?) Microsoft had to get something out of giving OpenAI millions of dollars in compute.

But Microsoft never wanted to be here. It had wanted to abandon Windows and Microsoft 365 for years, just to be pulled back in during the artificial intelligence boom due to the OpenAI investment. It felt pressure to get something out of the investment and capitalize on the boom itself, so it haphazardly reinvested in product, much to the chagrin of its users. Windows itself is underdeveloped, even by Microsoft’s own admission. (The company is finally remedying this after years of complaints.) Its core product — the piece of software it is known for — has languished in the face of frankly useless AI additions, like AI images in Notepad. (What?)

This was never how Microsoft drew up its plans, and now that OpenAI is building its own Stargate data centers with Oracle — if it can afford them — and Microsoft is released from its initial commitments, the company is easing back into being a model and inference provider. The whole Microsoft Build presentation on Tuesday signaled a shift away from consumer products and a concerted effort to appeal to developers. Microsoft has missed out on the OpenClaw phenomenon. It has kneecapped its GitHub Copilot plans due to high demand. Every major AI lab builds for macOS first and Windows when they remember it exists. And it’s not like normal consumers have given Microsoft a vote of confidence for its AI features. Clearly there’s room for a different strategy here.

I think new models are a great place to start. Microsoft is an incompetent product organization. Nobody there is driven by great products, unlike Apple or maybe even Google. Leave the Windows team to clean up the mess and start anew — build inexpensive, competent models and sell them to developers. The results will trickle down to users eventually in the form of better software, so long as agentic development is the future of software engineering. At a time when OpenAI seems close to collapse after its inability to market to enterprise users, Microsoft seems uniquely positioned to leverage its existing relationships — every technology company uses GitHub in some form — and massive cloud infrastructure. It doesn’t need to build data centers; it can subsidize inference with ease if it controls the models.

Microsoft’s divorce from OpenAI brings the potential to declutter Windows and bring plentiful inference to enterprise customers. I think that’s a good place for Microsoft to be.

Hopes and Ruminations on WWDC 2026

Eshu Marneedi — Thu, 04 Jun 2026 00:14:36 -0400

There are three things I want to cover in this post: the Snow Leopard rumor, the Siri shenanigans, and Private Cloud Compute. Beginning with the Snow Leopard rumor, here’s Mark Gurman, reporting for Bloomberg back in February:

Apple’s iOS 27 cleanup job includes code changes that could improve battery life. When it comes to Apple’s operating system updates this fall, the company has two main priorities. First, it wants to integrate AI into the software with a revamped, chatbot-style Siri. Second, the company is looking to clean up an operating system that has become a bit of a mess under the hood.

The tidying effort (similar to what Apple did with its Snow Leopard OS) will include removing scraps of old code, rewriting features, and subtly upgrading apps to let them perform better. That should make for a snappier, more responsive operating system. Apple also is planning some tweaks to the interface, though nothing as extensive as last year’s Liquid Glass introduction.

As part of the work on iOS 27 — code-named Rave within Apple — engineers are making improvements that could boost battery life, I’m told. The company hopes that underlying code changes will result in efficiency gains that will end up giving users more juice. If Apple does indeed pull off these improvements, it’s unclear if the company will market the changes — or if this is just a benefit of cleaning things up.

I think “Snow Leopard” is actually quite an apt description for this year’s operating systems. OS X 10.6 Snow Leopard was famously described as having “Zero New Features” onstage at 2009’s Worldwide Developers Conference. Indeed, the software had few customer-facing changes, with Apple instead opting to give the OS much-needed performance and stability improvements. Part of the underlying cause of these improvements, however, was an internal feature called Grand Central Dispatch, which introduced multithreading to OS X for the first time. Apple apps — and third-party apps later in the fall — were updated to run multiple processes concurrently, taking advantage of new Macs’ multiple cores.

Grand Central Dispatch is so important to modern iOS and macOS that we don’t even use the name anymore. Any processor-intensive task on iOS and macOS uses multithreading. (Just check Activity Monitor → Window → CPU History.) It was one of the most important features of OS X and set a precedent for how operating systems would work in the coming decades. Similarly, I imagine the new “more personalized Siri” and focus on on-device artificial intelligence — if these features are released into beta next week — will soon become integral to the way we use Apple devices. That was the thesis of my 2024 post-WWDC article “Why Apple Intelligence Is the Future of Apple Platforms.” On-device intelligence can repair AI’s suffering public image and takes full advantage of the powerful Neural Engines found in the latest Apple silicon.

But at its core, Apple’s platforms desperately crave some attention to detail. They’ve been deprived of engineering talent and design intuition for years, and it is starting to show. With the help of AI agents, I hope Apple revisits the SwiftUI apps written at the beginning of the decade and refines them with more capable UIKit and AppKit components. (System Settings, especially.) Conversely, I know there is plenty of old Objective-C code that could do with a modern rewrite — Swift is vastly more efficient than Objective-C, and with younger Apple engineers coming only with experience in Swift, now seems like a great time to begin the transition to newer components slowly. I understand the apprehension from some seasoned developers, but this transition must happen at some point. Whether it should include SwiftUI or not in the agentic age is debatable.

I also hope Apple’s new leadership lends greater importance to user experience. Alan Dye, Apple’s ex-software designer who left for Meta in December, famously neglected the Mac and its idiomatic controls and axioms. I hope Steve Lamay, Apple’s new software designer, pays greater attention to finer details, both on iOS and macOS. I hope new leadership, plus the unwavering spirit of AI agents, will have an important impact on the quality of Apple software come next week.

Bloomberg’s Gurman also has been extensively reporting on the “more personalized Siri.” Here’s him in May:

Apple has redesigned Siri for modern iPhone hardware, making it live inside the Dynamic Island as an always-on agent that can help users get things done across the operating system and within apps. The system can draw on web data, personal information, and what’s on a user’s screen to complete tasks.

There are two starting points for Siri. The classic approach — saying “Siri” or holding down the iPhone’s power button — is set to trigger a redesigned Siri animation in the Dynamic Island, the pill-shaped screen element that Apple introduced in 2022. That mode is best suited for voice-based queries and search.

The second method is entirely new: Apple plans to let users swipe down from the top center of the iPhone anywhere in the system to launch a new Search or Ask interface. (The Notification Center can now be opened by swiping down from the top left.) That opens a revamped Siri experience designed for getting things done or searching by typing, though using voice remains an option.

I’m quite intrigued to see how this interface looks in the final version of the OS, because Gurman’s reporting seems incoherent to me. Is the interface that appears when holding down the Side Button the same as the one hidden behind the Dynamic Island? What about iPhones without the Dynamic Island? What about the Mac and iPad? What differentiates the Siri mode from the Search and Ask interface? What is the difference between searching and asking? I think the answers to these questions will be a surprise come next week. Broadly, I think Apple has solved the three modalities: web search, device search, and agentic actions on the web. How it will integrate those three modalities in this interface is Monday’s question.

If I had to guess, I’d say the Side Button immediately launches the voice mode, while the Dynamic Island swipe gesture mirrors the current “Talk to Siri.” (I have Talk to Siri disabled because I find I activate it incidentally with the current double-tap-the-bottom gesture. This redesign will alleviate that pain.) “Ask” and “Siri” are different modalities for the Siri chatbot — text and voice, respectively — powered by on-device Apple Intelligence models and Gemini models in the cloud. The Ask feature will use a web search tool at its discretion. This is how I’d design the interface, at least. App Intents or Model Context Protocol integrations will likely be built into the Ask interface as tools, as will the Personal Context feature unveiled in 2024.

I’m also anxious to see how third-party providers function. I assume the plugin system Gurman has described in prior reporting is separate from the Ask/Siri interface. People will be able to talk to the Apple chatbot, which has all of their information and apps from the Personal Context and App Intents, but they’ll also be able to use ChatGPT instead if they prefer. It’s not like OpenAI will be a secondary model provider for the Siri interface — those models are bespoke, custom-designed Gemini models commissioned by Apple. Clearly the Siri chatbot will be the main affair, and secondary chatbots will be able to chat with users through the same interface.

I will note that I’ll find it odd that there’s even any doubt that Apple won’t ship these features in Beta 1. The company literally settled a class action lawsuit alleging that it misled customers about Apple Intelligence features in 2024 — the same features that still haven’t launched. In the post-Something Is Rotten in the State of Cupertino era, I think the world would completely lose it if Apple dared to come up onstage and say the new features are again coming “later this year.” It’s Beta 1 or bust — I don’t reckon anyone is willing to give Apple more timeline leniency.

Aaron Tilley, reporting for The Information:

As part of an Apple agreement with Google, some user queries to a new version of Siri will run in Google Cloud on a licensed version of the search giant’s Gemini model. Apple recently approved the use of a privacy technology from Nvidia in that setting, suggesting it will use Nvidia AI chips for at least some of its computing needs in Google Cloud, according to people familiar with the matter.

Specifically, Apple will tap into Google’s fleet of Nvidia’s Blackwell B200 data center chips, said the people. Apple will enable Nvidia’s confidential compute feature that encrypts data as it’s being processed on the chips.

The Information’s report also says that it is “unclear” how Apple will use Private Cloud Compute to run the white-label Gemini models, but I think the ambiguity is all the reporting we need. I contrarily predicted in January that Apple wouldn’t abandon Private Cloud Compute because it “wouldn’t squander the last remaining competitive advantage it has in the AI space,” and I stand by it. I don’t think Apple will hand all inference off to Google. While these Nvidia processors are fast and efficient at running inference, Apple has made lofty promises about the carbon neutrality and privacy of its AI products. This was a defining feature of the announcements two years ago. To think that Apple would abandon those virtues entirely at a time when backlash to AI is so outsized is foolish.

I’m not sure how Apple will split inference between Google’s servers and its own, but I imagine that the company will handle most inference either on-device or with Private Cloud Compute, not Nvidia and Google’s privacy stack. Maybe only the most difficult prompts will be sent to Google Cloud, or perhaps there’ll be an option to enable third-party inference in exchange for less privacy assurance. Something like “Allow Third-Party AI Processing.” Using Private Cloud Compute and touting 100 percent carbon neutrality during one of the most heavily watched technology presentations of the year would be an indescribably gargantuan boon to Apple’s AI efforts. If Apple came out and said, “Your data is confidential, we don’t train on it, it doesn’t kill the planet, and it doesn’t require a data center in your backyard,” I’d imagine that’d be one of the best AI pitches out of Silicon Valley.

Apple is undoubtedly using Google Cloud for training, since it has delegated that part to Google, but more efficient inference is what people seem to care the most about. Two years later, I hope Apple has figured this out.

Nvidia Announces a Consumer-Grade Processor at Computex

Eshu Marneedi — Mon, 01 Jun 2026 01:48:27 -0400

Sean Hollister, reporting for The Verge:

This fall, Nvidia will officially become a consumer PC chipmaker like Intel, AMD, Apple, and Qualcomm, putting a complete computing chip — not just graphics — into the very heart of laptops and mini-PCs. After many months of leaks, it’s finally announcing the RTX Spark, the first in a family of chips that will meet or beat the most powerful thin-and-light Windows machines ever, it claims.

“This is the most efficient PC chip ever built,” says Nvidia senior director of product management Mark Aevermann — without sharing so much as a single statistic or chart to back that up.

The RTX Spark is effectively the same GB10 chip that’s in the DGX Spark, the tiny “personal AI supercomputer” that Nvidia released last year, only now it’s a family of chips instead of just one. The flagship version appears to be spec-to-spec identical with 20 CPU cores, 6,144 GPU cores, and 128GB of LPDDR5X memory.

I find this announcement deeply fascinating. Not only is it scant on useful details, but it’s chock full of superlatives. This processor is somehow the “most efficient PC chip ever built,” which is such an audacious claim I don’t even know where to begin. What is a PC chip in this context? What is efficiency? Heat, power, price, all three? Are Macs “PCs?” Is Nvidia challenging Apple silicon’s industry-leading performance per watt? Nvidia seems loath to share any information, so we should take this all with a huge grain of salt.

I’m only linking to this because it illustrates how far gone the artificial intelligence hype cycle is at this point in 2026. Here’s a real quote from Hollister:

Nvidia claims this adds up to “a new personal computing paradigm where AI is the UX” and “users no longer need to master complicated app UIs” because you’ll just talk to your PC instead of needing to use mouse and keyboard.

None of this is intelligible. “AI” is not a user experience; AI is a technology that a user experience can be built around. ChatGPT is the user interface behind the GPT-5.5 large language model. The experience of using that UI is the “user experience,” by definition. This is a nonsensical statement, but I’d love to know what Nvidia means. Does it mean that chatbot-style user interfaces will replace all graphical windows and applications? That we will no longer have webpages to look at and windows to click on, replaced by chat bubbles? Or maybe the chat bubbles are gone entirely and we just speak to the sky and hope someone hears us.

I’m only jesting, of course. This is inscrutable marketing. Nvidia, by far the largest benefactor — and beneficiary — of the AI boom, wants to inflate its stock by positioning its new processor as at the forefront of computing. It wants its shareholders to believe that any computer without this chip is destined to fail because people will no longer want to use their hands and eyes to navigate a computer operating system, and that the RTX Spark is the most efficient — or perhaps only possible — way to meet these demands. It’s just a preposterous assertion. Transparently nonsensical.

UIs being complicated does not mean anything other than bad UI designers and computer scientists were put in charge of designing and crafting that UI. A “complicated” UI is not itself an indictment of the graphical user interface paradigm. Instead of highlighting how Nvidia’s chips can run the most powerful LLMs and how they can aid designers and programmers to make amazing UIs that people will intuit in seconds, the marketing people have taken over without even a moment of consideration. People like scrolling social media. They like typing documents. They like playing video games. Has anyone who wrote this presentation ever used a computer?

The answer to that, of course, is yes. They’re some of the brightest technologists in the world. But they’re betting their audience isn’t. And they might just be right.

A correction was made on June 2, 2026, at 5:05 p.m.: An earlier headline for this article said the RTX Spark was a laptop-only processor. This is incorrect. Microsoft unveiled a desktop PC at Microsoft Build that uses the RTX Spark processor. I regret the error.

Why Is Anthropic Making Money?

Eshu Marneedi — Thu, 28 May 2026 16:20:26 -0400

Mike Isaac and Cade Metz, reporting for The New York Times:

Anthropic, once the lesser-known artificial intelligence competitor to OpenAI, has been on an inexorable rise over the past few months.

The San Francisco company recently dueled with the Pentagon over the use of A.I. in warfare. It released a powerful A.I. model, Mythos, that it said was uncannily capable of finding and exploiting hidden flaws in software. And the company advised Pope Leo XIV on his papal encyclical delivered on Monday, in which he warned about safeguarding humanity from A.I.’s most disruptive effects.

On Thursday, Anthropic punctuated its ascent by officially passing OpenAI as the world’s highest flying A.I. start-up. Anthropic said it had raised $65 billion in financing that values it at $900 billion before the inclusion of the new capital, a deal that puts it ahead of OpenAI’s last valuation of $730 billion. The company also unveiled a new flagship A.I. model, Claude Opus 4.8, which is significantly better than its predecessor at generating computer code.

$900 billion on $47 billion of revenue is hysterical. The AI industry is very much in a bubble, and both Anthropic and OpenAI’s valuations will deflate significantly in the months and years after their initial public offerings, rumored for later this year. Exactly when that deflation happens is debatable, but it definitely will. Anthropic is almost a trillion-dollar company — a distinction it took Apple decades to achieve as the first to hit the milestone in 2018 — and it does not even make $100 billion in revenue. For context, Apple’s revenue run rate upon reaching a $1 trillion market capitalization was over $255 billion.

But Anthropic is unique as the only AI laboratory to turn a profit, due to its massively profitable enterprise deals and stingy rate limits. Anthropic does not like subsidizing heavy usage for its customers, as I wrote about in April. In the bubble, Anthropic might be the only properly valued company. OpenAI is struggling financially — it is unable to pay for its massive data center projects and has no financially successful product for an IPO. In a vacuum, it is OpenAI that is overvalued, not Anthropic.

So why does Anthropic “deserve” a $900 billion valuation? I’ve been thinking a lot about this as OpenAI rapidly iterates its Claude Code competitor, Codex. The OpenAI team has been shipping updates faster than any tech company I’ve seen, and every update is quite impressive. The Codex remote control feature, allowing users to code from their iPhones, is fully featured and excellent. The Computer Use plugin is astounding. The Codex plan rate limits are the best in the industry, and people can even use them with third-party tools like OpenCode and T3 Code. By comparison, Anthropic’s products are substandard. The Mac app is almost unusable, the Dispatch remote control feature is basic, and the computer use tool is nothing like OpenAI’s.

OpenAI is objectively winning the product development race. Codex is an immeasurably better product than Claude Code, and people on social media adore it. I hear almost nothing about Claude Code anymore among the developer crowd, in stark contrast to just a few months ago. Yet everyone is loath to cancel their Claude Code subscriptions — certainly enterprise customers, sans Microsoft, which is transitioning its employees to its bespoke GitHub Copilot product. While Claude doesn’t meaningfully surpass OpenAI’s models in any benchmarks — they appear to be neck and neck — the experience of using the model is absolutely better. This has created an interesting debate between models and products. OpenAI has the best products and great models, whereas Anthropic has fantastic models and terrible products.

This chasm will only widen in the coming weeks as Anthropic deploys Mythos-level models to the public. (Claude Mythos is Anthropic’s name for a much more intelligent model that surpasses the latest OpenAI and Google products.) Anthropic’s models have such an intuitive sense for what makes for great code, which is especially important to enterprise customers. While vibe-coders on social media care least about code quality, major corporations want their codebases to be workable. Fundamentally, Codex is better at working through problems and writing many lines of code to fix them, whereas Claude Code is slow, methodical, and detail-oriented. They’re just different mindsets entirely. Only one appeals to people who look at code. The Anthropic models are better at writing code — the OpenAI models are better at solving problems.

Anthropic will continue to be profitable as long as the models are more important than the products.

At I/O, Google Hopes to Ditch the Humans

Eshu Marneedi — Thu, 21 May 2026 15:54:18 -0400

Apps, harnesses, and the singularity

Who asked for this? Image: Google.

I think Google has landed on a solution to the shortcomings of Google Search: Gemini.

During Google’s I/O developer conference on Tuesday, where the company announced updates to its products, Gemini and artificial intelligence once again took center stage. The company, live from its headquarters in Mountain View, California, released a smarter yet smaller large language model, Gemini 3.5 Flash, to better position Gemini in Google Search, its flagship product used by billions of people daily. It previewed new programming tools for developers, including a refreshed version of Antigravity, Google’s AI coding agent powered by Gemini. And it used that new coding harness to power a consumer-facing agent built into the Gemini app, called Gemini Spark, a system that runs in the cloud with access to a full computer to complete tasks autonomously on the web.

These are just a handful of announcements Google showcased during its Tuesday presentation, but I think they’re the most important because they illustrate an overarching strategy. Google wants to get people into Gemini, whether they’re programmers, casual Gemini app users, or Google Search users. At I/O this year, the company has fully committed to Gemini as the future of its search and generative artificial intelligence endeavors, and now hopes its waffling for the past few years hasn’t undermined the strategy. Google, more than any Silicon Valley technology company, has both users and ecosystem lock-in. Its job now is to capitalize on those users, provide a compelling product, and hope people stick with it.

Google has so far accomplished none of those objectives, but at least it now has a coherent product strategy. Gone is the haphazard mishmash of contrasting products — Gemini, AI Mode, Gemini Agent, Project Mariner, Google Assistant, Gemini on Android, the Gemini command line interface, and the plethora of other consumer projects Google has released since the start of the AI boom in 2023. The company’s core product strategy revolves around Gemini — the Gemini app, that is — and it took the first steps on Tuesday to funnel the billions of Google Search users to Gemini over the coming years.

Product coherence aside, I don’t find the mission all that compelling. The new Antigravity harness is lackluster and leaves much to be desired compared to Claude Code and Codex, its main competitors. The new model is also smaller than the other flagships, packing in fewer parameters, or “intelligence.” AI Mode in Google Search, in my eyes, is a complete failure and distraction from the inaccurate, sloppy AI Overviews that have littered Google Search in recent years. The Gemini app, including Gemini Spark, seems impressive, but is it worth axing web search for? And Google still has not made clear how publishers can capitalize on the boom. With the company’s strategy defined by Gemini, links to external websites will only become sparser. What happens to the open web as a result of these changes? What happens to humans as a result of these changes?

There are many adjectives one could use to describe Google I/O this year: boring, nonchalant, disappointing. I’d say the best one, though, is clarifying. But the lack of ambiguity here might not be comforting.

Search

Google positioned this year’s updates to Google Search as the largest since the search engine’s launch over 25 years ago. I agree, not because the announcements were that significant, but because the core search product hasn’t changed much since its inception. At the center of Google Search, now and for the past two and a half decades, is a list of blue links to websites. Over time, what has appeared at the top of and beside those results has varied. Initially, it was nothing — just the links, and nothing more. As Google grew and made money, those spaces became home to lucrative advertisements, some of the best places to promote a product or business on the entire internet. As Google’s machine learning models developed and the Knowledge Graph grew, so-called “featured snippets” occupied the top of the page, perhaps with images and other visual aids.

Today, at the top of almost every Google Search page is an AI Overview, generated by one of Google’s Gemini Flash models. Beginning Tuesday, Gemini 3.5 Flash, the latest model, will generate these overviews. AI Overviews replace featured snippets and often prioritize speed over correctness. Initially, they used inferior models to deliver quick responses — now, they use a better model but with a low, or zero, thinking budget. The model is encouraged to deliver a fast response without checking its work. The result is that AI Overviews are quantitatively poor at both generating correct answers and summarizing work on the internet. A New York Times investigation into Google’s AI Overviews found that the latest version, running on Gemini 3.1 Flash, produced completely fabricated answers 9 percent of the time. Google’s own research pegged the number much higher, at 28 percent.

More worryingly, the overviews’ answers tend not to be grounded in sources, increasing the possibility of hallucinations when asked questions that require fresher knowledge. The Times investigation reported that, when using Gemini 3.1 Flash, Google’s AI Overviews did not provide a source for 56 percent of correct answers. This is almost certainly not a user interface glitch — the model relies heavily on information stored in its own parameters rather than summarizing the information found in search results, as the latter takes more time. The model must open each of the links, store the content in its context window, then summarize it. The Times investigation further stated that the AI Overview occasionally identifies a “reliable website but seems to misinterpret its information.” At other times, the AI Overview relies on Facebook or Reddit posts, especially when reputable content is sparse.

This is the primary experience most people have with generative artificial intelligence. Google Search is used by billions of people daily, and Gemini is used by around 900 million, per Google. AI Overviews are, by every heuristic measure, much less accurate than even Gemini 3.1 Flash in the Gemini app harness, let alone frontier models such as Gemini 3.1 Pro. Google is astutely aware of the limitations of such overviews, so it released AI Mode alongside AI Overviews last spring. AI Mode appears to act like Gemini with enhanced web search capabilities. For one, it always searches the web, whereas the Gemini app does not. It also offers the larger models, like Gemini 3.1 Pro, even to free users. AI Mode is accessible via the AI Overview by simply tapping on the summary — doing so will generate more content, albeit more slowly. Users can also ask follow-up questions from this mode.

Google said Tuesday that around one billion people use AI Mode monthly, a statistic I view with skepticism. I believe Google is counting people who tap on AI Overviews for more information as AI Mode users, which is technically true but disingenuous, because tapping on the summary does not regenerate the response for accuracy — it simply adds more information. Anecdotally, I have yet to hear from a single person outside tech circles who uses AI Mode regularly by clicking the AI Mode button in Google Search. People who are likely to reach for an AI-generated response, as opposed to passively consuming a summary, tend to use Gemini instead. I don’t doubt that 900 million people use Gemini monthly — the app appears on par with ChatGPT in popularity.

All of this is why I tend to believe the AI features in Google Search are, at best, a bandage for a self-inflicted problem and, at worst, an abysmal failure. For one, AI Overviews are objectively incorrect more than any other AI product, including the ever-popular ChatGPT free tier. They give large language models an unnecessarily bad reputation, and they give people incorrect information. That does not further Google’s stated mission of “organizing the world’s information and making it useful.” And AI Mode, buried under a button or summary, still is not nearly as accurate or useful as the Gemini app itself. Neither of these features promotes the overall health of the open web, a dying sector of the technology industry whose concerns Google continues to dismiss. One day, we will live in a world where publishers will stop publishing and the vast majority of content will be written by LLMs.

We are already seeing the effects of the content vacuum today — Google’s AI Overviews refuse to link to publishers because there are no publishers to link to. And if there are no publishers, there’s no training data. Google’s AI search endeavors have refused to solve the fundamental problem with Google Search: poor results stemming from search engine-optimized filth. Instead, they are merely accelerating the problem’s metastasis, discouraging the publication of reliable information online. They’re inaccurate, and they make their training data inaccurate because there isn’t enough good data to train on anymore. There aren’t enough good sources to link to. Google Search is heading toward an infinite death spiral.

Again, Google is astutely aware of this, which is why it must transition search away from links. Google does not want people to visit other websites — it wants them to interact with its chatbot, Gemini, to produce more training data. This is a short-term solution to a long-term plague, but it is what Google has ultimately decided on. It hopes to never have to rely on publishers because Gemini will become so good one day that it creates the data for itself. Theoretically, with this approach, Gemini will not have to rely on Reddit threads for advice — it will be “smart” enough to reason through a problem by itself and come up with an answer. Google wants to cut the publisher out of the scenario entirely. If publishers want a part, they can advertise to have their products or links favored within Gemini chats.

Google keeps referring to this idea by saying people are typing “longer, more detailed queries” into Google Search. That might be true, but I also think it’s at least a bit of an exaggeration. Google wants people to be frustrated by the “10 blue links,” subconsciously begin treating Google Search like Gemini, then transition those users to AI Mode first and Gemini later. Once they’re in Gemini’s territory, using smarter models with more comprehensive reasoning traces, the model can act as an agent, scouring much older, more niche information for an answer. It is no longer forced to rely on Reddit threads or social media posts, the links most likely to float to the top of search results. It will no longer need a publisher to report the news. The agent becomes the journalist or the DIYer or the YouTuber sharing niche knowledge. This might seem like an absurd idea, but it’s very much where Google believes the internet is heading. Google wants the internet to be by agents and for agents — it does not want people to do the searching or the writing anymore. It is killing their traffic, and it is making publishers’ livelihoods impossible.

Framing Google’s Tuesday announcements in this light makes the overarching strategy abundantly clear. The Google Search bar, beginning Tuesday, will autocomplete queries to be longer and more sentence-like, similar in structure to a Gemini prompt. Autocomplete suggestions come from a bygone era, when PageRank, the algorithm Google Search uses to rank pages by their authenticity and relevance, used keywords and links to find appropriate websites. This is how SEO became such a vibrant industry — littering a website with the most commonly suggested keywords was a great way to get to the top of the page. Now, Gemini doesn’t rely on keywords — it needs a lengthy, detailed query to reason well.

When people enter the results, Google will automatically route them to AI Mode whenever it deems necessary to get people accustomed to waiting a while for responses. Google Search results are effectively instantaneous because nothing is generated. AI Overviews have struggled to meet this standard, but AI Mode has no such pressure for brevity or speed. AI Mode functions much like Gemini, and again, the end goal is to encourage people to use the mainline Gemini app and agents generally. AI Mode this year borrows many features from Gemini, including the Antigravity harness to write code and put together mini applets to explain concepts, for instance. It also includes an agent similar to Project Mariner that can interact with other websites autonomously. The point is to change people’s ideas about Search — it is no longer a place for people to find information.

This is what Google meant by “the biggest changes to Google Search.” Google is done with the 10 blue links, but it is also done with publishers, social media, and websites made by humans for humans. The open web, it hopes, is a thing of the past, to be replaced by agents that interact with each other. It does not even want people to shop online by themselves anymore — agents can instead navigate sites and purchase goods through the Universal Commerce Protocol. Google wants people to spawn agents across its products, including YouTube, AI Mode, and Gemini, to purchase products without user confirmation. Gemini will hunt for products based on user-set criteria, ensure each specification is met, and automatically place the order with a retailer. Project Mariner laid the foundation a year ago, and now, taking heavy inspiration from the Model Context Protocol, Google hopes people will trust their agents with their credit cards.

And that is where it becomes apparent that hardly anyone wants this. People want information on the open web; people do not want to give an agent their credit card. Fans of Google — who have reached out to me this week expressing their displeasure at my recalcitrance toward agentic search — point to the one billion people who “use” AI Mode and the 900 million who use Gemini. But those people don’t have a choice — Google has systematically destroyed the internet and web search and sold people the solution: AI Mode. Agentic products are useful on many occasions, but the rise of AI Mode sharply correlates with the unmistakable decline of Google Search. AI Overviews are objectively unhelpful. People, anecdotally, are frustrated with the quality of AI-generated responses on Google and the sheer number of unhelpful results cluttering the links. Traffic to publishers has dropped steeply while the demand for that content only seems to rise.

AI Mode is a failure; it is a symptom of a dying web. It is a solution to a problem Google created. This year’s I/O positions that solution as a bridge between the Google Search of yore and Gemini-powered agents that Google hopes to capitalize on.

Gemini and the Singularity

The most feature-rich of the Gemini-powered agents thus far — and the first to graduate from preview status — is Gemini Spark, an instance of Gemini 3.5 Flash running in the cloud via the rewritten Antigravity harness. Gemini Spark is Google’s answer to OpenClaw, and the first always-on agent that combines Personal Intelligence, computer use, and apps through the Model Context Protocol. Project Mariner, introduced in 2024, manifested itself in Gemini Agent, a computer-use tool that planted the seeds for an always-on agent that could perform tasks autonomously. Gemini Spark is a continuation of this endeavor, relying on the newer Gemini models’ improved tool-use capabilities to remain in standby forever. The agent is effectively given its own computer in the cloud — people can prompt it like Gemini Agent, but it can also work autonomously.

That’s almost all we know about Gemini Spark. The service is “coming soon” for Google AI Ultra subscribers. It will supposedly have access to all of a person’s Google services: Gmail, Google Calendar, Google Drive, and more. It will only support Google tools at launch, but the agent will be compatible with third-party MCP servers and, supposedly, eventually any other website on the web, similar to the Gemini Agent feature it now succeeds. It will have access to a command line via the Antigravity command line interface, allowing it to view URLs; download and edit files using the myriad command line utilities available for agents; and compile code. Gemini Spark is a more feature-rich, cohesive alternative to OpenClaw, the open-source, vibe-coded agent written by Peter Steinberger, now an OpenAI employee.

Google is once again betting that agents such as OpenClaw and Gemini Spark are the future of human-computer interaction. There might be some validity to that belief, but I think it’ll require a seismic shift in how people think about their computers. Currently, the personal computer is a place where people go to get work done. Google Search is the “library,” Google Drive is the “desk” with the “paper” and “pens.” Gmail is the mailbox. Google — and broadly, the rest of Silicon Valley — hopes that one day, the computer itself manages the library, the desk, and the mailbox; that there will be no need for a human to check their email or search the web or write a document by themselves. And that’s not even considering how the Valley sees software code in the mid-2020s.

Demis Hassabis, the chief executive of Google DeepMind, the company’s AI laboratory, closed out the I/O presentation on Tuesday by saying that the company believes humanity is “at the foothills of the singularity,” referring to a point in the creation of generative AI at which humanity undergoes an irreversible change in lifestyle. It has become increasingly evident over the past year that the “singularity” the Valley refers to — and the “artificial general intelligence” that will presumably bring humanity to it — means a broad rearchitecting of the role of humans in science, the arts, and the information sector. The Valley, let alone Google, does not have an answer to what that role may be post-singularity. It doesn’t even seem all that perturbed that the central question around the role of humanity has gone unanswered for over three years.

The only objective that is clear is that Google wants people to use Gemini so aggressively that it can charge a fortune for ads and subscriptions. It wants people to get so accustomed to AI that they’re willing to pay hundreds of dollars for the privilege of using it. It wants normal people to forget about the open web and consume hyper-personalized ads in the middle of their chatbot responses. Google wants to capitalize on the AI boom through advertising and subscriptions. (Google has all but forgotten about the enterprise market, a point I’ll come to shortly.) Whatever happens after this presumed increase in revenue — which, I should add, is yet to come close to transpiring — is everyone else’s problem. Google just needs people to be alive and rich enough to spend money. It, for now at least, has no interest in the philosophical considerations of the technology it has so nonchalantly developed year after year.

Keeping this product-oriented philosophy in mind helps clarify the true motivation behind Google’s latest product announcements. Gemini Omni, the company’s new “omnimodel” — a transformer model that generates video and images — is better at producing animated diagrams and instructional videos. Google doesn’t care what happens to instructional YouTubers who have made billions of dollars on its platform, YouTube. Neither does it care that generative AI has decreased literacy rates and caused a decline in English test scores. It is unperturbed by the fact that the rise of YouTube in the classroom is a distraction. It just hopes Gemini Omni cuts the humans out of the equation cheaply and efficiently, to the point where the model becomes a drug for lazy creators or teachers. The same is true for the newly announced text-to-music model Lyria 3 Pro. None of these models “inspire creativity.”

None of this is to say Gemini is not a useful product, which 900 million people might dispute. But this year’s I/O has shown that Google sees Gemini as the eventual successor to its human-crafted products. Google Search now funnels people into Gemini; Gmail and Google Workspace cede territory to Gemini almost every month; even YouTube is not immune to Gemini. For as long as the AI bubble lasts, Gemini will drive Google’s product strategy — what happens to humans is a concern for another time, post-“singularity.”

Antigravity

Behind all of Google’s Tuesday announcements was a reworked, rewritten agentic coding harness. A harness is a set of tools and instructions given to a model to let it work autonomously. ChatGPT is a harness around GPT-5.5 — Codex, Claude Code, and the Gemini app are also harnesses. AI Mode, too, is a harness around Gemini 3.5 Flash, giving it access to a web search tool, thinking traces, and tools to help lay out responses. This is different from a user interface, but the distinction is more subtle. People can use the ChatGPT harness in the ChatGPT app and on the website, for instance. The Codex harness is available in the Codex app and the command line interface. Different user interfaces, same harness.

Gemini entered the agentic coding scene with the Gemini CLI, an open-source harness and command line interface that competed with the then-newly launched Claude Code. That harness, despite Gemini models excelling at nearly all coding benchmarks, was historically quite poor. Claude Code wasn’t great, either, but Claude models, beginning with Claude Opus 4.5, became so good at using tools that it didn’t matter. The Cursor harness had always been the best way to use Gemini models.

Google also introduced a second harness after the launch of Gemini 3 in December, called Antigravity. Antigravity was a competitor to Cursor, the AI-enabled Visual Studio Code fork, created after Google acquired Windsurf, itself a Cursor competitor. It was a full code editor modeled after VS Code and supported the latest Gemini models in a new harness. This harness was better than the Gemini CLI but was ultimately still spiritless, and Antigravity instead became known for its generous rate limits for Claude Opus 4.5. Until Tuesday, Antigravity was the graphical user interface for using Gemini models for coding, while the Gemini CLI remained the command line harness.

On Tuesday, Google rewrote the Antigravity harness and released both a GUI and CLI to improve Gemini 3.5 Flash’s tool-calling abilities. The new harness is simply named Antigravity, and the two user interfaces are Antigravity 2.0 and Antigravity CLI. The former is simply a graphical equivalent to the latter, written with Electron and taking heavy inspiration from Codex and Claude Code. The harness itself is rewritten in Go and powers agentic features in AI Mode and the Gemini app. When tested, Gemini 3.5 Flash using the new Antigravity harness outpaced Gemini 3.1 Pro in every software engineering benchmark, according to Google’s results. Artificial Analysis, an independent company that runs a suite of benchmarks to produce a single “intelligence” coefficient, puts Gemini 3.5 Flash without the new harness considerably behind Gemini 3.1 Pro and GPT-5.4.

It has become obvious why Google’s AI efforts have not turned a profit: the models are still poor at coding by themselves, leaving a lot of pressure on the harness. Enterprise customers spend billions of dollars on tokens for coding, and Anthropic has been able to capitalize on that market due to its models being the most proficient at tool calling, even if those models are more expensive to run. Theo Browne, a YouTuber and programmer who tests LLMs, said in a video that Gemini 3.5 Flash isn’t considerably better than prior Google models, and I tend to agree with that analysis. The model, while good for knowledge work like any other Google model, is still created with an old pre-training run from many years ago. All Google models since around Gemini 2.5 have only been freshly post-trained; their weights are still out of date.

At the risk of turning this article into a model review, I think Google, in order for its AI bet to be successful, must accelerate the development of its models for programming. Coding is the most lucrative market for LLMs and Google simply has no answer to the billion-dollar contracts both Anthropic and OpenAI have won in the last few months. And I have a tough time believing these are the models that will lead us to the “singularity.”

There was a litany more announced at I/O this year that I have no interest in covering. There are thousands of articles examining every new feature coming to all of Google’s products; the firehouse of new AI products is exhausting to write about. But I think 2026 should be a year to step back and observe the progress the industry has made so far. Three years in, AI favorability is at an all-time low, and Google’s newly clarified product direction does not appear likely to reverse that trend. If the Valley truly wants to climb the mountains of the singularity, perhaps it should develop an answer to what happens to humans — publishers, users, creators — at the summit.

Google Announces the ‘Googlebook’ and Gemini Intelligence

Eshu Marneedi — Wed, 13 May 2026 08:48:21 -0400

Antonio G. Di Benedetto, reporting for The Verge:

Google is announcing a new line of laptops coming in the fall called Googlebooks. Details are sparse for now, as the tease is just a small part of various Android announcements during Google’s Android Show. But we do know this is a major new initiative in the laptop space for Google, seemingly designed to succeed Chromebooks with something more capable: a platform running a long-rumored new operating system based on a fusion of Android and ChromeOS…

So, what do we actually know about Googlebooks and their operating system that’s not Aluminium but also not not Aluminium? For starters, Googlebooks are built on the Android technology stack. They’ll run Chrome for web browsing and also run Android apps. They’ll even be able to directly access files from your Android phone and run your apps right off of it so you don’t have to temporarily move your attention across devices. And they’re going to have Gemini Intelligence baked into just about everything — right down to the cursor.

Googlebooks will have a Magic Pointer feature that offers contextual suggestions whenever you shake your cursor and point it at something on the screen. Google’s examples include setting up a meeting by pointing at a date in an email or selecting images of furniture and a living space to visualize them together. Beyond your mouse pointer, Googlebooks will also feature the custom AI-created widgets that Google is also debuting today for Android phones and Wear OS smartwatches.

“Googlebook” is indeed an asinine name, but I was half-expecting it to be called “Geminibook.” Nevertheless, the Googlebook platform is built with “Gemini Intelligence,” a suite of Gemini features that aim to do agentic work in apps people have installed on their devices. If this sounds familiar, it is. Apple announced the same feature, powered by App Intents, almost two years ago. It is yet to ship, of course, and Google has a track record — through Project Mariner and Gemini Agent — of shipping features similar to Gemini Intelligence. From Allison Johnson:

Task automation is apparently one of those “best of Gemini” features. It’s already on some recent Pixel and Samsung Galaxy phones, and it enables Gemini to use certain apps on your behalf. It’s been limited to a handful of rideshare and food delivery apps until now. That’s changing “soon,” says Google, when task automation will open up to a wider range of apps.

It will also add multimodality; previously, Gemini could only use voice or text prompts to inform its actions. Now you can throw a screengrab or a photo into the mix, which kind of seems like something you should have been able to do from the start. You’ll be able to give Gemini a screenshot of a grocery list in your notes app and it will add those items to your cart. You know, provided you have an Android phone that supports Gemini Intelligence.

The task automation feature is intriguing because, while Google probably wants to use the Model Context Protocol and application programming interfaces to control apps long-term, it has, in the short term, settled on a computer-use agent trained to tap around apps. Think of Codex’s computer use or the browser-use agent in ChatGPT Atlas, OpenAI’s web browser. This is also more or less how Project Mariner worked, mostly because the MCP servers tech firms are publishing — if they make them at all — are limited in scope. That’s why ChatGPT’s “Apps” feature has failed to reach velocity — it just isn’t useful. (I have said many times that computer-use agents and MCP-powered apps are highly inefficient, and I stand by this. APIs, or now command-line interfaces, will be the future of this kind of work. App Intents are just APIs.)

There is utility to an agent that can work in other apps, but it depends on the implementation. I don’t see the usefulness of an agent that can order DoorDash or Starbucks. Such implementations are useless because people usually prefer to choose what they want to eat. (I realize this might be surprising to the Silicon Valley types.) I also don’t think people are in any hurry to give an agent their credit card information so it can spend thousands of dollars on their behalf booking flights or hotels. As Brian Chesky, the chief executive of Airbnb, said on social media, chatbot-style agents are the wrong user interface for these choices. There is little wrong with an intake form to make granular choices.

Back to the Googlebook, a “platform” for Google’s next-generation, unnamed desktop operating system based on Android. The Chromebook platform is bespoke, built on Google Chrome, which has limited the variety of apps that can run on ChromeOS. The Googlebook is not a hardware laptop model — it is an operating system and a list of specifications for third-party hardware manufacturers. It will nominally succeed the Chromebook, but in practice, Chromebooks are mainly used by schools that use them because they’re based on Chrome. Years of Chromebook usage in elementary and middle schools across the United States have made this generation practically tech-illiterate. They don’t know what a file system is, they don’t know how applications work, and they abstain from using the keyboard and trackpad. It’s far too late to transition them to a “real” desktop operating system — it’s an operational nightmare for the Chromebook’s most popular demographic.

So that makes the Googlebook a competitor to the MacBook Neo, which is regrettably preposterous. The MacBook Neo’s singular strength is that it’s a great Mac. It runs full-fledged macOS with all the programs people love, and attention to detail and beauty. It’s Windows, but better. And if people aren’t interested in an alternate operating system, they’ll purchase a cheap Windows laptop with no learning curve and all of the programs they’re used to. The market for the Googlebook is unclear. I don’t think anyone loves Gemini enough to switch their entire computing life to a blown-up version of Android just to shake their cursor and have Gemini summarize whatever is on screen. I can see why Google wants to make a desktop OS — I support the endeavor! — but that ship has, unfortunately, sailed long ago.

I said earlier on Tuesday that the Android Show this year was exciting, and I maintain my enthusiasm. There’s a lot of new stuff here, especially Gemini Intelligence. But, much like all of Google’s products, I’m afraid they won’t have much staying power.

OpenAI and LoveFrom’s Hardware Device Might Be a Smartphone

Eshu Marneedi — Mon, 11 May 2026 06:50:40 -0400

Tim Hardwick, reporting for MacRumors:

OpenAI is said to be fast-tracking development of its first “AI agent phone,” with the company now aiming to mass produce the device as early as the first half of next year, according to industry analyst Ming-Chi Kuo.

Late last month, Kuo revealed OpenAI’s work on a smartphone, contradicting earlier reports that the company had no plans to enter the mobile market. Kuo said MediaTek and Qualcomm are the chosen chip partners and Luxshare Precision Industry is the exclusive manufacturing partner, with mass production scheduled for 2028.

I was quite bullish on Jony Ive’s pedigree of making beloved products, so much so that I initially expressed optimism about the collaboration between his LoveFrom design firm and OpenAI, but it appears that was wrongheaded.

The ChatGPT app, for three years, has remained at the top of the App Store and Google Play Store. It is one of the most popular mobile applications of all time. When people want to talk to ChatGPT, they do so from the app and website. Frequent users add Home Screen and Lock Screen widgets for easy access, and power users can invoke ChatGPT by pressing Option-Space on their Mac if they have the desktop or Codex apps installed. ChatGPT is practically everywhere — the only market the company is missing out on is the Apple Watch. What I’m saying here is that there is zero friction to using ChatGPT on any device today. It might even become more frictionless after the “more personalized Siri” presumably launches in the fall.

People, under no circumstance, will ever purchase a smartphone made by OpenAI. This market is entirely nonexistent. If these rumors are credible, and OpenAI is indeed making a smartphone — either running its own operating system or, most likely, Android — it will become one of the most embarrassing failures in consumer technology. Even disregarding the lack of friction in using ChatGPT on iOS or Android today, generative artificial intelligence is one of the most disliked technologies in the world. A smartphone littered with “AI slop” — and that will be almost entirely vibe-coded — will be widely regarded as an abysmal experience. Nobody wants this. Nobody wants to pay OpenAI for anything, let alone an AI slop phone that’s worse than their iPhone and uses one of the most disliked technologies in existence.

Of course, this rumor is coming from Ming-Chi Kuo, who leaks at the supply chain level. This could be a very different product that utilizes the same processors and other components. I hope it is — I don’t think Ive is so myopic that he’d choose to build an Android smartphone. That doesn’t sound like Jony Ive, and the only way OpenAI could ever get him to make something like that is if they paid him more money than the sun is worth. Such a device is anathema to the core ideas Ive has espoused throughout his career at Apple: simplicity, utility, and beauty. Generative AI doesn’t make mobile user interfaces simpler, more useful, or more beautiful. Agents have great potential in knowledge work, but they’re no good at scrolling social media, watching video, or capturing photos. This is what people do on their smartphones, and if they ever want to conjure up an agent, well, the ChatGPT app is right there.

What are we even doing anymore?