General evaluations

Frequent testing when scaling model or dataset

Testing models after significant increases in compute, data, or model parameters. Even relatively small changes to model or dataset size can introduce new properties (“emergent abilities”) and failure modes. Identifying them early can prevent the models from being released prematurely [9, 22].

Using an AI model to evaluate AI model outputs

In cases where the outputs of AI models cannot be easily evaluated, AI models can be used to evaluate their outputs or the outputs of other AI models [82, 16, 17, 91]. The evaluations can then provide a training signal to improve the original model’s performance or offer explanations of the output for human users.

GPAI models explaining model outputs in a zero-sum debate game

Debate is a technique that aims to produce reliable explanations of AI model outputs that are too complicated for humans to understand, by letting two GPAI models role-playing in a debate produce an explanation in a dialogue [102].