GPT-3 model learns to express uncertainty in natural language (openai.com)

Summary:

• A GPT-3 model can generate answers and express its confidence levels about its own answers in natural language.
• The model remains calibrated under distribution shift and is sensitive to uncertainty in its own answers.