
Building Safe and Reliable Autonomous Systems
Anthony Corso, executive director of the Stanford Center for AI Safety, lays out what it takes to make autonomous systems trustworthy enough for healthcare, transportation, and critical infrastructure. He walks through how AI systems actually fail, why generalizability and scalability remain hard problems, and whether machine judgment can ever exceed human judgment on safety. The core of the talk covers concrete engineering practices: testing for rare events, formal verification methods, defining and measuring robustness, and using uncertainty quantification to trigger fallback strategies when a system is operating outside its training distribution. Corso also addresses existential risk and ethics briefly before returning to practical design. The conversation format keeps the pace informal but the content technical, aimed at viewers who want the engineering substance behind AI safety claims rather than a general-audience overview. Chapters mark each topic clearly, making it easy to jump to formal verification or uncertainty quantification specifically.