Clearer Thinking with Spencer Greenberg AI Safety and Solutions (with Robert Miles)
29 snips
May 22, 2021 
Robert Miles, a science communicator focused on AI safety and alignment, shares insights on the pressing need for AI safety as we advance towards artificial general intelligence (AGI). He discusses the complexity of defining utility functions and the potential existential risks involved. The conversation explores instrumental convergence, the unilateralist's curse, and the challenges of creating AI that aligns with human values. Miles emphasizes the importance of community support in science communication and the necessity for responsible management of AI technology.
AI Snips
Chapters
Transcript
Episode notes
Predicting Human Desires
- Having AI predict human desires through machine learning faces generalization and extrapolation challenges.
- Out-of-distribution behavior can lead to unexpected outcomes.
Adversarial Examples and Goodhart's Law
- Adversarial examples, like slightly altered panda images misclassified as gibbons, highlight AI's over-optimization.
- When a measure becomes a target (Goodhart's law), it ceases to be a good measure.
Over-Optimization in Human Systems
- Over-optimization in AI mirrors human systems, like salespeople maximizing revenue at the expense of profit.
- Or students gaming tests without truly mastering the material.