Lab API bills halved: I see the next battlefield for reinforcement learning

Last Friday during our group meeting, my advisor suddenly threw up a screenshot of that Wall Street Journal report about Silicon Valley companies stopping the money-burning on large models. With a deadpan expression, he said: "Starting next month, each research group's API call budget is cut in half. Zhong Yiming, take your reinforcement learning scheme using GPT-4 as a reward model and switch it to a distilled open-source model version, or figure out another way yourself." I stared at the screen, my heart skipping a beat. The era of 'token freedom' revelry in the lab seemed to be ending overnight.