Watching a language model think before it speaks

A language model's thought process is visualized in real-time using a new interpretability technique called the Jacobian lens, allowing users to see the model's internal workings and correct its mistakes before it generates an answer.