“OpenAI recommends using Reinforcement from Finetuning (RFT) for problems where no existing model on the market can perform the required task, particularly in verifiable domains like chip design and biology.”