Unlike current goal-seeking agents, the 'Scientist AI' is designed as a 'pure predictor' without inherent goals or preferences about the world's state. This non-agentic foundation is intended to provide mathematical guarantees of honesty and safety, avoiding the pitfalls of instrumental goals and reward hacking.
Impact: High. By decoupling prediction from agency, Bengio aims to create a fundamentally safer AI that does not pursue its own objectives, thereby mitigating existential risks.
In the source video, this keypoint occurs from 00:11:46 to 00:14:02.
Sources in support: Yoshua Bengio (Guest, AI Researcher)

