Spotlight: Jul 23, 2024
Humans decide which tasks to use general-purpose large language models for, “so we have to take the human in the loop into account,” says Ashesh Rambachan. A new method evaluates a model based on its alignment with a human’s beliefs about its capabilities.