Apple Intelligence launches this autumn, limited to iPhone 15 Pro, which excludes 80 to 90 percent of the existing iPhone install base. No one outside Apple has used it yet. The foundation models Apple published benchmarks for are competitive with current market offerings, but the architecture is the real story: Apple built an LLM with no chatbot. You cannot send a raw prompt and receive a raw response. Emails get prioritized, summaries appear on button press, Siri returns GUI instead of model output. The LLM is abstracted as an API call.
This is a direct strategic bet against AI maximalism, the thesis that general-purpose chatbots will replace discrete software by handling complex multi-stage tasks through a single prompt interface. Apple's counter-thesis is three-part: generative AI is most useful when embedded in a system with existing user context, when unbundled into specific features running as small efficient on-device models, and when it carries zero marginal cost per inference at mass-market scale. That third point is not rhetorical. Nuclear power station economics do not survive consumer pricing expectations.
The full piece earns a read not for its conclusion but for the structural argument it builds around Apple's developer incentives, the commodity status it assigns to ChatGPT, and the question of who actually controls the context layer in a world where the OS owns the user data. Benedict Evans has been skeptical of the software-is-dead camp in print before, and this is his most detailed technical and strategic case yet for why the unbundled, embedded model beats the maximalist one.
[READ ORIGINAL →]