New Ways to Corrupt LLMs

This characteristic is particularly useful to add to one's mental model of how context and prompts interact with token clusters (anthropomorphized as concept synthesis or generalization).