The marketing name is "on-device intelligence" or "private AI"; the filing name is adaptive training at the deployment destination. Apple's granted patent US11636348B1, issued April 25, 2023 and naming inventors including Ruslan Salakhutdinov, claims adapting neural-network models where they run — on the device — rather than centralizing all learning in the cloud. Its CPC classes G06N 3/088 (unsupervised learning) and G06N 3/08 (neural-network learning) place it in core ML training art.
On the record, the privacy story is a consequence of the architecture, not the other way around. If a model adapts on the device, the personal data used to adapt it never has to leave the device. Apple markets that as privacy; engineering-wise it is a decision to push training, not just inference, to the edge. The patent claims the harder half of that decision — adapting the model in the field.
Novel, or just renamed? On-device inference is now common; on-device adaptation — letting the model learn from your data locally — is the more ambitious step, and it is genuinely harder because training is far more compute- and memory-hungry than inference. The claim's interest is in making that adaptation work within a device's constraints. The principle of model adaptation is established; doing it at the deployment destination on consumer hardware is the claimed work.
The strategic frame is that on-device AI is also a margin and services story. Inference and adaptation that run locally reduce server cost and create a privacy differentiator Apple can sell against cloud-first competitors. The patent is the technical anchor for a competitive position the marketing makes about values.
Scope, stated carefully: this is a granted patent (B1 kind code) on a specific adaptive-training method, within an enormous edge-ML landscape that Google, Qualcomm, and Samsung also crowd. It evidences Apple's architecture, not a monopoly on on-device learning.
Follow the filing, not the keynote. When Apple says its AI is private because it runs on your device, the precise claim behind the slogan is something like this 2023 grant: adapting the model where it is deployed. Privacy is the marketing; deployment-destination training is the record.