Apple researchers taught a model to write working SwiftUI code even though the training data had almost no Swift examples. They ran cycles of generation, compilation checks, and GPT‑4V filtering to produce nearly one million valid programs and a stronger model called UICoder. The result shows self‑improvement can work but raises worries about reliability and scaling of synthetic datasets.
ORIGINAL LINK: https://www.techradar.com/pro/apple-employees-produced-an-llm-that-taught-itself-to-produce-good-user-interface-code-independently
