Apple released a toolkit for fine-tuning its Foundation Model that powers Apple Intelligence features, allowing developers to customize the model's behavior for specific tasks. The author experimented with creating a Golden Gate Bridge-obsessed AI assistant by training an adapter with custom data. Key findings include the model's preference for short responses, the need for powerful hardware (32-64GB RAM Mac or cloud GPU), and the ability to modify behavior despite Apple's safety constraints. The toolkit uses adapter-based fine-tuning, requiring only 160MB downloads instead of full model weights.
64 Impressions