You can keep raw user text off public trainers. You cannot promise that a synthetic row has no echo of a real person if someone pasted a chat into the seed. Review is the control. There is no magic filter.
DeviceSignals only
DeviceSignals only
ConsentOn-device or your Mac
DeviceSignals only
ConsentOn-device or your Mac
StopPublic trainer upload
Limits you should write in the README
- Default telemetry has no transcripts.
- Text leaves the device only after a clear yes, and only to systems you name.
- Synthetic data can miss real phrasing and can still leak if the seed leaked.
- A LoRA can memorize a rare string. Do not train on secrets.
- A new
.aimodelis an app release, not a silent update of Apple's system model. - Evaluations and Instruments traces can hold prompts. Treat them as sensitive.
If a regulator, a contract, or your gut says the seed is too close to a user, drop the row. A smaller honest set beats a clever flywheel.
If the failure is tone, try a prompt version first. Version prompts like product copy is cheaper than a bake. If the failure is a tool path, fix the tool and the eval. Train when the model must carry a style or a domain the system model will not hold.
Primary sources: Core AI, apple/coreai-models, and Evaluations.
Short form. Signals first. Synthetic rows with review. LoRA on a Mac. Bake to Core AI. No raw chats to a public trainer by default.
Key concepts
- Default telemetry has no transcripts.
- Text leaves the device only after a clear yes to systems you name.
- A LoRA can memorize a rare string. Do not train on secrets.
- A new
.aimodelis an app release, not a silent system model update.
Takeaways
- Signals first, synthetic rows with review, LoRA on a Mac, bake to Core AI.
- Drop rows when the seed is too close to a user.
- Fix tool paths and evals before training when the failure is not model knowledge.