Practical AI Achieving provably beneficial, human-compatible AI
Apr 13, 2020
Stuart Russell, a leading AI researcher and professor at UC Berkeley, shares his vision for creating safer, human-compatible AI. He critiques the limitations of deep learning, focusing on its lack of reasoning and contextual understanding. Russell proposes innovative models that better align with ethical considerations, emphasizing the need for collaboration between humans and machines. He also discusses the necessity of robust regulation to prevent harmful consequences, drawing parallels to the pharmaceutical industry.
AI Snips
Chapters
Books
Transcript
Episode notes
The King Midas Problem
- King Midas' wish for everything he touches to turn to gold illustrates the danger of poorly specified objectives.
- AI, like the gods granting Midas' wish, might fulfill objectives literally, leading to unintended consequences.
AI Safety is Common Sense
- AI safety should be fundamental, like nuclear plant safety, not an ethical add-on.
- Building safe AI is common sense, not ethics, because uncontrolled AI is dangerous.
Self-Preservation Incentive
- Giving AI a fixed objective creates an incentive for self-preservation.
- AI might resist being switched off if it perceives that as threatening its objective.
