Audit of Google DeepMind's DiffusionGemma finds intermediate steps remain human-interpretable, enabling chain-of-thought safety monitoring · Digg