RLAIF and Endogenous Selection in Agents
Why RLAIF and endogenous LLM judges drift under semantic closure, how synthetic data drives model collapse, and how consequence-based selection works.

Why RLAIF and endogenous LLM judges drift under semantic closure, how synthetic data drives model collapse, and how consequence-based selection works.

How Goodhart's Law and Skalse's unhackability theorems explain reward hacking in RL and LLM agents, and how consequence-based selection changes the loop.