LessWrong (Curated & Popular)

“Limitations on Formal Verification for AI Safety ” by Andrew Dickson

Aug 27, 2024
Andrew Dickson, an expert in formal verification and AI safety, dives deep into the challenges of ensuring AI reliability. He discusses the limitations of formal verification in messy real-world scenarios, where full symbolic rule sets often fall short. The conversation highlights the complexities of predictive modeling in biology and the difficulties in simulating human interactions. Dickson emphasizes the ongoing need for rigorous inspections, arguing that even with advancements in AI, achieving strong guarantees remains a daunting task.
Ask episode
Chapters
Transcript
Episode notes