(Time as an API. Formerly known as spreadsheet. Can any old app be reincarnated as prompts? What are the top questions of the day? How does crypto correlate to the markets? Dartmouth or DARPA? )

Long contexts represent a new front in the struggle to control LLMs. We explored a family of attacks that are newly feasible due to longer context lengths, as well as candidate mitigations. We found that the effectiveness of attacks, and of in-context learning more generally, could be characterized by simple power laws. This provides a richer source of feedback for mitigating long-context attacks than the standard approach of measuring frequency of success