OpenAI's 2023-12-14 experiments used models already trained on broad data. Supervision means additional training examples or feedback; here a weaker model supplied it to a more capable model. The question was whether this could draw out the stronger model's existing capabilities without merely copying the supervisor's mistakes. Some methods exceeded the weak supervisor on tested tasks. The authors call this a proof of concept: training on ChatGPT preference data (judgments about preferred answers) failed, and the model-supervision setup differs from human supervision. Weak and strong refer to relevant capabilities, not size alone. The proposed analogy to humans supervising more capable AI does not establish reliable human review of unfamiliar scientific discoveries.

Evidence status: bounded_empirical_analogy.