Comment by HPsquared
1 year ago
This is the Waluigi effect.
Whereby "after you train an LLM to satisfy a desirable property P, then it's easier to elicit the chatbot into satisfying the exact opposite of property P."
1 year ago
This is the Waluigi effect.
Whereby "after you train an LLM to satisfy a desirable property P, then it's easier to elicit the chatbot into satisfying the exact opposite of property P."
See also: [[Streissand Effect]]