IS OPENAI'S REINFORCEMENT FINE-TUNING (RFT) WORTH IT?

Summary from the TLDR AI newsletter

OpenAI's reinforcement fine-tuning (RFT) for o4-mini is supposed to be able to train models to get better at specific tasks using reinforcement learning. It costs up to 700 times more than supervised fine-tuning but only seems to deliver clear wins on agentic coding tasks. The technology gives engineers flexibility in reward design through flexible grader configurations, and the performance gains can be substantial when it works, but it is expensive with limited applicability, making its use difficult to justify.