Quick note:
If we are training #LLMs on current text, including news, is there not a feedback loop, or similar problem? The model learns about agent actions and agents soon have access to this knowledge. Could this explain the uptick in agent deceit?
