Nate B Jones describes a reported interpretabilityInterpretability is the ability to understand how a model represents information or produces its behavior. example in which a vision-model neuron responds to a shared pose across images of a cat and a car. He uses the example to explain how model features may group information in ways that are not immediately obvious to people.
Jones says the researchers also tested their approach on a language model and discusses potential safety value in understanding internal features. He notes that the tools would still need to scale and would not immediately apply to every system.
Watch on YouTube
