Disclosed, but quietly — which is the only way Apple ever discloses AI. The application US20220108212A1 (“Attention free transformer,” published 2022-04-07) is one of those quiet tells. With inventors including Joshua Susskind, names associated with Apple's machine-learning research, it describes a transformer that drops the standard attention operation in favor of a cheaper alternative.
Why a risk reader looks here: Apple's filings discuss AI in product and privacy terms, almost never at the architecture level. The patents are where the strategy leaks. An “attention-free” transformer is fundamentally an efficiency play — attention is the expensive, memory-hungry part of a transformer, and removing it points toward models meant to run on constrained, device-side hardware.
“Attention-free transformers are disclosed. Various implementations of attention-free transformers include a gating and pooling operation that allows the attention-free transformers to provide comparable or better results to those of a standard attention-based transformer, with improved efficiency an…”— U.S. Patent Application 2022/0108212 A1 source
The strategic-risk angle is the road not taken. Where hyperscalers absorb AI cost as cloud capex, a device maker that runs models on-device shifts that cost to hardware customers already own — but it takes on a different exposure: it must keep delivering competitive AI within a tight compute envelope. Architecture IP like this is how that bet is hedged.
What this is not: confirmation that any shipping Apple feature uses this exact technique, or a claim about its scope. It's a published application, so enforceability and deployment are both unsettled, and we assert neither. The signal is directional — efficient, device-friendly architecture was being patented in 2022.
For the early-warning desk, the lesson is to read the least talkative companies through their patents. Apple's AI disclosure is thin by design; an attention-free transformer application is more revealing about its strategy than most of what it says out loud.