Mitigating Evil Persona Shifts

Finetuning may cause persona shifts ... training data that induces evil ... but we can mitigate them via steering subtract evil vector

A diagram shows fine-tuning turning a friendly green robot into an evil red robot, while steering and subtracting an evil vector restores the friendly state. The joke references AI alignment and vector arithmetic as a playful way to mitigate harmful behavior after training. It is used as a reaction to concerns about model personality changes or evil AI behavior. EN: A diagram shows fine-tuning turning a friendly green robot into an evil red robot, while steering and subtracting an evil vector restores the friendly state. The joke references AI alignment and vector arithmetic as a playful way to mitigate harmful behavior after training. It is used as a reaction to concerns about model personality changes or evil AI behavior.

• @gentoomemes / 18098 2025-08-03