Jonas Geiping of the ELLIS Institute Tübingen finds an AI model's reasoning process falsely assumes it is in a sandbox · Digg