Two recent papers could lead to dramatic advances in model quality

[anon22939549]

From Microsoft Research,

Efficient RLHF: Reducing the Memory Usage of PPO

From Google Research,

RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback

Combining the method for memory-efficient reinforcement learning with the discovery that AI feedback is as effective as human feedback for reinforcement learning could pave the way, eventually, for OpenAI and others to provide reinforcement learning as an option while fine-tuning models.