Learning Foundation Models and PCC
·1 min read
aifoundation-modelslearningapple
Apple's Foundation Models framework and Private Cloud Compute are still new enough that most developers haven't built with them yet. Here's how I'd structure the learning path.
Phase 1: Core ML fundamentals
Before touching Foundation Models, get comfortable with Core ML:
- Create ML — train a simple text classifier or image classifier
- Core ML Tools — convert a PyTorch model to
.mlmodel - Understand model size vs. accuracy tradeoffs on device
- Profile with Instruments to see actual inference time
Phase 2: Foundation Models framework
Once Core ML clicks, move to the Foundation Models framework:
- Set up a
LanguageModelSession - Experiment with
@Generableschemas for structured output - Build a simple tool-calling flow with
@Tool - Test with the Evaluations framework
Phase 3: Private Cloud Compute
PCC extends the on-device model to Apple silicon servers:
- Understand the trust model — no data retention, verifiable by researchers
- Learn when the system automatically routes to PCC vs. stays local
- Test your prompts across both tiers to see where quality differs
Phase 4: Production patterns
The real work is in production:
- Token budgeting — on-device models have small context windows
- Graceful degradation — what happens when the model is unavailable?
- Evaluation loops — use
ModelJudgeEvaluatorto track quality over time - Privacy auditing — verify no user data leaks to unintended endpoints
Resources
- Apple Foundation Models documentation
- WWDC sessions on on-device AI (2025, 2026)
- The
BookTrackerandWritingToolssample projects from Apple