How to assess a general-purpose AI model's reliability before it's deployed

A new technique estimates the reliability of a self-supervised foundation model, like those that power ChatGPT, without the need to know what task that model will be deployed on later.