Mystery AI Hype Theater 3000 Episode 1: "Can Machines Learn To Behave?" Part 1, August 31, 2022
11 snips
May 31, 2023 AI Snips
Chapters
Transcript
Episode notes
AI Value Alignment's Narrow Focus
- AI value alignment discussions often avoid deeper questions about AI's role and potential harms.
- Focusing on aligning AI with "our" values ignores power dynamics and whose values are prioritized.
Toxicity and Power Dynamics
- Questions like "Who gets to define toxic data?" often dodge existing power imbalances.
- Discussions about toxicity in datasets lack the nuance of other frameworks addressing discrimination.
Modeling Minds? Not Really
- Blaise Aguera y Arcas claims AI can "model people's states of mind," citing an archive paper.
- The cited research merely trains a model to classify pre-annotated text, not infer mental states.