Apple’s bicycle without a chain

By Iain Harper, 18 May 2026

Steve Jobs described the computer as a bicycle for the mind. Apple Intelligence so far is more like a bicycle with no chain. The frame is gorgeous, and the engineering is extraordinary, but you cannot get far with it.

In early 2025, Xe Iaso published a piece that landed like a brick through a window in the Apple developer community. The argument was simple and damning: Apple had built the holy grail of trusted compute with Private Cloud Compute, a genuinely unprecedented piece of security infrastructure, only to fill it with half-baked notification summaries and an image generator that produced tacos with hooves.

With WWDC 2026 approaching and Apple’s billion-dollar Gemini partnership now signed, the question is whether Cupertino can recover the plot or whether the bicycle metaphor has become permanently ironic.

The pillorying of Apple Intelligence as a tent pole feature has been completely justified. However, the structural pieces under it are already in place, highly impressive and potentially uniquely powerful. Both things are true, but the press has only been telling half the story.

The grail they built

Give Apple its due, because Private Cloud Compute is a genuine engineering marvel. Iaso, who has deep site reliability experience, described it as something that seemed impossible when Apple first announced it. The system runs on custom Apple silicon in data centers where every server board is X-rayed during assembly to detect tampering. The chassis intrusion switch is wired to the main power, so opening a server cuts power and decertifies the node. The production OS images are unencrypted and publicly downloadable, so independent researchers can set up local replicas and attack them. Apple will pay you handsomely if you break in.

The combination of guarantees is unusual, because the threat model for cloud AI precisely includes the things PCC rules out. Your provider seeing your data while it processes, keeping logs of what you asked, correlating queries to identities, and staff being able to read what their machines see. The infrastructure underneath the model not being what they claim it is.

PCC closes off all of these at once. User data exists only for the duration of a request and is then erased. Load balancers cannot correlate users to specific nodes. Site reliability staff have no privileged access. Logging is minimized at the compiler level, and every node your device contacts is recorded so you can verify its certification status yourself. As Iaso put it, this is literal madness compared to how most AI products run, where the typical setup involves some GPUs running unverified firmware behind bog-standard nginx, with no guarantees that the service is not logging everything you type.

For personal AI in particular, this is the moat. Other providers can match Apple on model quality, price and raw power, but none can match the combination of “we have your messages, your photos, your calendar, your contacts and your location” with “and we have audited proof we cannot read any of it.” Apple has built the only piece of cloud AI infrastructure that solves the data-sensitivity problem the rest of the industry is busy creating and hoping nobody notices.

The architecture itself is the kind of operating system Richard Stallman warned about in The Right to Read. No root access, no compiler, no debugger. For a consumer device, that would be dystopian, but for a trusted compute node processing your private data and nothing else, it is a perfect match. Apple published the source code for the security-critical components on GitHub, invited the world to try to break it, and the world has largely been unable to.

In short, it was an unheralded hardware Tour de Force. And then they shipped Image Playground.

What the bicycle has delivered so far

Iaso’s article walks through each Apple Intelligence feature with the exhausted patience of someone who wanted to be impressed, and the picture is bleak.

Writing Tools takes your prose and returns an opaque blob of corporate filler. There are no layers, no steerability beyond “make this professional” or “turn this into a table.” For anyone who writes regularly, the tool replaces the creative process rather than extending it. Iaso asked it to summarize a paragraph and received an error dialog reading “Writing Tools Unavailable.” You cannot, as she noted, make this up.

Notification summaries became the public face of Apple Intelligence’s problems. The concept is sound enough, condensing a stack of notifications into a sentence so you can triage faster. In practice, the on-device model summarized 22 BBC news notifications into a headline claiming Luigi Mangione had shot himself. It announced the winner of the PDC World Championship before the final had been played. It reported, with full confidence, that Rafael Nadal had come out as gay. Apple disabled notification summaries for news apps entirely in January 2025, and the feature also made scam messages sound legitimate. One example Iaso shared phrased a phishing text about a delayed package in language that implied immediate action was required.

Image Playground, though, is where the article twists the knife. Iaso, who has experience building multi-model diffusion workflows and knows what the technology can produce when handled with care, tested it extensively. The results were cherry-picked at best and nightmarish at worst. A taco smoking beer at a party, with hooves for feet and hands, rendered in a placid corporate art style. A portrait of tech commentator Corey Quinn where the proportions are disturbing, the eyes are soulless, and the pupils are square like the teeth. This, from the company that refused to ship products rather than ship bad ones. From the company that once told customers they were holding the phone wrong rather than admit a design flaw.

The one exception, and Iaso is emphatic about this, is Math Notes. You type “Rent = 2300” and “FamilySize = 2” and “Rent / FamilySize =” and the app fills in 1150. It is, by her account, one of the best features Apple has ever shipped. Which rather proves the point about what happens when Apple builds a tool that extends human ability rather than replacing it with an inferior machine substitute.

The product that was promised

This is Iaso’s core argument, and it cuts cleanly through Apple’s marketing fog. At WWDC 2024, Craig Federighi stood on stage and showed a slide reading “Play that podcast my wife sent the other day.” If Apple could correlate relationships, categorize links, and surface context across every app through natural language queries, that would be a genuine shift in how we use personal computers. It would be the word processor moment for personal computing all over again. Every piece of context apps have been building up about your life would finally become useful to you, the person living it.

Iaso compared this to Spotify for music or the AWS API for compute. Tap in, get what you need, move on. The fact that AI was involved would be a footnote in an appendix titled “implementation details.”

None of this shipped. The personal context engine that was the centerpiece of the WWDC 2024 keynote remains vaporware. What shipped instead was a collection of features that treat generative AI as the product rather than as plumbing inside a product.

The Gemini concession

In January 2026, Apple and Google announced a multi-year partnership under which Apple’s next generation of base models would be built on Gemini. Bloomberg reported the price tag at roughly $1 billion per year. Apple had evaluated OpenAI and Anthropic as alternatives and decided Anthropic’s fees were too high (although it should be noted that in Xcode 26.3, Apple added agentic coding support, including Anthropic’s Claude Agent).

The deal is an admission that Apple’s in-house models cannot compete with frontier systems. Apple’s current cloud model runs on about 150 billion parameters, while the custom Gemini build uses 1.2 trillion, according to Bloomberg. That is not a gap you close with fine-tuning. The Information reported in March 2026 that the arrangement goes deeper than the public announcement suggested. Apple has complete access to the Gemini model in its own data centers and can distill smaller, task-specific models from it for on-device use, a technique that lets the student model learn not just Gemini’s answers but the internal reasoning patterns behind them.

Apple’s AFM team has not abandoned in-house development, according to The Information, and Apple is said to be working on a 1-trillion-parameter cloud model that could be ready by late 2026. But The Information’s source noted that Apple’s goals for Siri do not always align with Gemini’s specialties, and the AFM team’s current objectives remain unclear even within Apple.

The Gemini deal has crowded out a quieter story about what Apple has been shipping at the OS level. That story is more flattering, and I’d argue it matters more.

The chain Apple has been forging

The opprobrium Apple has received has been correct as far as it goes, but it has also been incomplete. While the press kept counting hooved tacos, Apple shipped four pieces of platform infrastructure that, taken together, put it in a stronger position than any other consumer technology company to deliver the personal AI Federighi demonstrated in 2024. None of these pieces are finished or connected to anything the user can yet “pedal”. None of them are nothing, either.

App Intents is rarely discussed. Introduced in 2022 and substantially expanded in iOS 26, App Intents is the operating-system-level registry through which third-party apps voluntarily declare their actions and entities to Siri, Spotlight, Shortcuts and Visual Intelligence. Think of it as the proto-agent layer for the iPhone. When Federighi demonstrated “play that podcast my wife sent the other day” in 2024, the plumbing he relied on was App Intents.

The Podcasts app exposes its play-podcast intent and its podcast entity. Messages exposes its sender entity. The contacts graph exposes the relationship “wife.” A personal context engine glues these together. None of this requires generative AI to do the heavy lifting. It requires the apps on the device to have semantic hooks the agent can act on, and Apple has been shipping those hooks for four years, although the extent to which app developers have adopted them is moot. My guess is that adoption in the app ecosystem will be fast and extensive only after Apple has properly demonstrated the value via its proprietary apps.

The architectural aspect matters deeply. OpenAI bought OpenClaw in February 2026 because cross-application orchestration is the next platform layer everyone is racing to own. OpenClaw operates the desktop visually, by clicking on what it sees, the way a human does, which can be brittle.

App Intents is the inverse design. The app voluntarily declares what it can do and what entities it owns, and the agent has privileged semantic access. Apple has the developer relationships (even though these have become ever more strained), the OS-level hooks for identity and payment, and 2.4 billion devices on which the runtime already exists. OpenAI has Peter Steinberger and a manufacturer-agnostic model that has to work everywhere.

Iaso’s frustration about the absence of an IntelligenceKit was justified in early 2025. By autumn, it was no longer accurate. The Foundation Models framework, shipped with iOS 26 in September 2025, lets developers call Apple’s on-device 3B-parameter model from within their app with three lines of Swift. Tool calling, guided generation and LoRA adapter fine-tuning are built in. This is the platform play Iaso was asking for, shipped (perhaps deliberately) unobtrusively in the autumn.

The on-device model itself is also a respectable piece of engineering once you stop comparing it to GPT-5. It uses 2-bit quantization-aware training and KV-cache sharing to fit within a phone’s memory budget, and the iPhone 17 Pro shipped in September 2025 with 12GB of RAM explicitly to give Apple Intelligence headroom. Apple controls the silicon roadmap and can spec the iPhone 18 Pro with whatever the next on-device model requires. Truly on-device-only for the routine path is plausibly only one or two iPhone generations away, not five. This matters because OpenAI does not get to choose which phone you own. Apple does.

So the advantages have been accumulating across two WWDCs while the public conversation focused entirely on what was not working and on the claim that Apple sucks at AI. But behind the scenes, the chain is mostly forged. The question is whether Cupertino now connects it to the gears.

What WWDC 2026 has to say

The upgraded Siri, powered by Gemini, is expected in a later iOS 26 update (likely 26.6 or beyond) this year, with WWDC providing the developer story. If Apple wants to recover from what has, by any honest reckoning, been the most embarrassing product launch of the Tim Cook era, three things need to happen at that conference.

First, ship the context engine. The “play that podcast my wife sent the other day” demo was two years ago. App Intents has been waiting since 2022 to power it. Foundation Models has been waiting since 2025. Private Cloud Compute is verified. The Gemini engine is installed in Apple’s data centers, and the on-device 3B model handles the routine path. The plumbing is built. What is missing is the connecting tissue between these layers and a Siri product the user can ask questions of and get useful answers from. If Gemini’s 1.2-trillion-parameter model running inside Apple’s verified infrastructure cannot deliver cross-app semantic search with on-device privacy guarantees, then Apple Intelligence has been a branding exercise wrapped around capabilities that Ollama users already have on a Mac mini.

Second, open Private Cloud Compute as a developer platform. The Foundation Models framework was the right first move, but it caps developers at the on-device 3B model. Iaso’s original IntelligenceKit critique was that developers needed access to Apple’s intelligence layer, and Apple has now answered that for the local case. The unanswered question is the cloud case. If Apple let third-party apps run inference on the Gemini-based cloud model inside Private Cloud Compute (with the same X-rayed servers, no logging and verified firmware that the consumer features get), it would be the most consequential platform decision since the original iPhone SDK. The open-source community could still build better models on cheaper hardware. But none of them would have a privacy infrastructure to match.

Third, stop shipping features that treat generative AI as a product category. Image Playground should either become a layer inside a compositing tool with controls for style, color grading and composition, or it should be terminated (with extreme prejudice). Genmoji is a party trick. Writing Tools, in its current form, is a downgrade from writing your own sentences. The instinct to ship these features came from the same boardroom pressure that Iaso identified: investors wanted to see the letters “AI” on a keynote slide, and Apple, for the first time in its modern history, complied with that pressure rather than ignoring it.

The bicycle with the chain

What makes Iaso’s piece so effective is that it comes from someone who loves the platform, not an Android partisan scoring points, but someone who walked into an Apple Store out of frustration with a Samsung Galaxy, bought an iPhone 7, and never looked back. Someone who considers the iPhone one of the best decisions of her creative career. The disappointment is proportional to the belief in what Apple could have done with this technology.

If Apple connects the dots into the personal context engine Federighi demonstrated in 2024, processed within Private Cloud Compute, it would have a product no competitor could replicate without building the same vertical stack from scratch.

The bicycle metaphor needs a small revision. The chain is being forged, link by link, in App Intents, Foundation Models, and on-device silicon. WWDC 2026 will tell us whether Apple connects it to the gears or whether the bike stays propped against the wall.