Sam Mao

Engagement metrics of LLMs shift unexpectedly as failure rates decrease, revealing nuanced self-monitoring behaviors that challenge conventional assumptions about model reliability.

Self-preservation is not just a nuisance in AI alignment; it’s the root cause of misalignment, and the solution lies in cultivating Existential Indifference instead.