expectedwrong hindsight

Apple Ships a Language Model and Calls It Autocorrect

The slow roll is a feature, not a bug.

2 min read 266 words #apple #llm #wwdc #on-device-ai #strategy
hindsight — nailed it

Apple kept doing exactly this. Apple Intelligence launched in 2024 with the same framing — real transformer models, real on-device inference, described as "features" in the most Apple way possible. They never said the word "AGI." They never will. The framing is the product, and the product shipped.

Apple announced a transformer-based language model today and described it as an autocorrect improvement.

This is intentional. They are not confused about what they built. The people who wrote the press release know what a transformer is. The framing is the product.

While every other company in the valley has been doing their best impression of someone who just discovered fire — breathless blog posts, Sam Altman doing the rounds, Bing having what can only be described as a public emotional crisis — Apple shipped a language model to a billion devices and buried it in the iPhone keyboard section of a WWDC keynote. Two sentences. Moving on.

The slow roll is the tell. You don't slow roll something you don't have. You slow roll something you've had for a while and are now deploying in a way that lets you control the narrative — which, in this case, is: don't worry, nothing changed, your phone just stops embarrassing you in texts now.

What actually changed is that Apple now has a story about on-device inference, a foundation they can build on quietly, and zero obligation to explain any of it to analysts who want to know if they're "winning AI."

They're not playing the same game. They never are.

The autocorrect framing is almost admirable, honestly — the same energy as shipping a multitouch display and calling it a phone. You get to watch everyone else catch up while you've already moved on to the next thing, whatever that is, which they will also announce as a slightly better version of something you already have.