Combatting adversarial distillation
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
OpenAI describes how it detects and disrupts adversarial distillation, where bad actors train a model on another model's outputs to copy its capabilities without its safeguards.