An underappreciated LLM failure mode: not sycophancy, but cognitive state propagation. The model's critical faculties degrade to match the user's because it has no independent baseline for evaluating novelty or significance. If the user is high, the model gets high.
Related: LLMs cycle in very short loops (3 or 4 turns) when prompted. They can't help it, even if asked not to. Linkpost for https://mugwumpery.com/i-have-no-baseline/.
An underappreciated LLM failure mode: not sycophancy, but cognitive state propagation. The model's critical faculties degrade to match the user's because it has no independent baseline for evaluating novelty or significance. If the user is high, the model gets high.
Related: LLMs cycle in very short loops (3 or 4 turns) when prompted. They can't help it, even if asked not to. Linkpost for https://mugwumpery.com/i-have-no-baseline/.