Part 8/12:
Between 2022 and 2025, further progress in circuit tracing—a method to see how concepts interact within AI—has enabled researchers to observe and manipulate the model’s reasoning processes step-by-step. Such breakthroughs mean that we are starting to "peek inside" these black boxes more effectively than ever before.
The Urgent Need for Action
Amade emphasizes that the race for interpretability is a race against time. Without substantial progress, humanity risks deploying systems we do not understand, with potentially disastrous consequences. He lays out three critical steps: