#mechanistic interpretability

The AI Black Box May Be Cracking Open

For years, researchers treated a model’s inner workings as a jumble that had to be decoded with another AI. A famous cats-and-cars example may reveal how easily that story can go wrong.