Base Models Beat Aligned Models at Randomness and Creativity

Peter West, Christopher Potts

Alignment has quickly become a default ingredient in LLM development, with techniques such as reinforcement learning from human feedback making models act safely, follow instructio…