“Brendan Foody is optimistic about Reinforcement Fine-Tuning (RFT) because it is profoundly data-efficient, requiring only hundreds to thousands of examples, which makes it viable for application-layer companies to customize models.”