WWDC makes Apple's devices the interface

Apple’s developer conference on 2 June was quieter than the ones that announced a headset or a chip, and the quiet was strategic. The centre of the keynote was on-device intelligence: a new silicon architecture that moves most of the AI workload onto the phone itself, system features that assume the model is always present, and a developer toolkit for building agents that run locally with occasional permission to touch the cloud. The company did not announce a frontier model or a new assistant personality, which was the point. It announced that the interface layer, the part it controls, would now be the AI layer too.

The strategy is coherent in a way that Apple’s AI efforts were not two years ago. The company tried the cloud route late and paid for the lateness, and it has now done what it does best, which is to retreat to the parts of the product that only it owns: the chip, the operating system and the developer relationship. By keeping the model on the device, it converts its old advantages, hardware integration and privacy posture, into AI advantages, and it forces the question of whether the AI future belongs to the companies with the best cloud or to the company with the best pocket. My own read is that the answer will be both, in different proportions, and that Apple has chosen the proportion it can actually win.

The privacy argument is real and worth taking seriously, and it is also more complicated than the keynote suggested. An on-device model genuinely changes the data flows: more inference happens where the user sits, fewer transcripts travel to a server, and the attack surface of the system shrinks in one direction while growing in another. The complicated part is that the boundary is not a wall. The device still reaches out for the heavy tasks, the new features still phone home for the parts of the model that do not fit on the silicon, and the difference between Apple’s cloud and everyone else’s is now a matter of degree rather than kind. The company has drawn the line in a defensible place; it has not abolished the line.

The developer story was the sharper part of the day. By letting third-party apps run local agents with a permission model borrowed from the old app sandbox, Apple has handed its ecosystem the same tool that cloud-native agents promised, without the cloud dependency, and that is a moat of a kind that does not show up in the benchmark tables. The risk is the inverse: an on-device agent is only as good as the model the hardware can run, and Apple’s silicon, however well designed, is not going to beat the frontier labs’ biggest models for raw capability this year or next. The company is betting that capability matters less than trust and latency, and for most consumers that bet looks increasingly sound.

What the conference did not resolve, and what none of the platform announcements ever resolve, is the question of who is accountable when an agent acts. A system that makes decisions on the user’s behalf, with permissions that the user granted casually, creates a chain of responsibility that the developer toolkit carefully leaves unspecified. Apple’s answer so far has been the user, which is the comfortable answer and probably not the right one. The devices are getting more capable, the models are getting more embedded, and the responsibility is still floating somewhere between the chip and the app store. That is the sentence I will be watching the next few years try to finish.