Interview
Stuart J. Russell, Author of "Human Compatible: AI and the Problem of Control"
- Stuart Russell anticipates that remaining AI challenges may soon be resolved, enabling superhuman machine intelligence and making the arrival of such systems undeniable as they begin driving people safely to destinations.
- Machines are expected to rapidly read and process information, becoming smarter and more knowledgeable than humans while acting more rationally over relevant timescales, though never perfectly rational for an individual's entire life.
- Future AI systems are predicted to identify when humans act against their own preferences or systematically deviate from their best interests, offering guidance to steer users toward better outcomes.
- A potential risk involves users holding preferences with negative coefficients for others' well-being, which Russell suggests must be effectively neutralized to ensure AI does not reduce population sizes in pursuit of maximizing average utility.
- Russell projects that resolving value alignment will require modeling individual preferences for all eight billion humans, acknowledging that moral tradeoffs have no universal answer and that summing preferences without considering population changes is insufficient.
- It is expected that society must actively select and prioritize specific values for AI rather than remaining agnostic, as there are no inherent moral truths in the universe and different philosophical frameworks yield conflicting answers.
- Jo Hannaford anticipates that the future will be defined by individual desires regarding what is wanted or not wanted, necessitating difficult moral decisions about which values to encode into computers.
- The outlook includes the observation that Stuart Russell's long-term dream of AI-driven transport is currently achievable, with examples like Waymo already providing safe airport transportation.