Time To Say Goodbye Has AI Sold Out? with Jessica Dai and Ben Recht
5 snips
Aug 20, 2025 Jessica Dai, a UC Berkeley PhD student and writer, and Ben Recht, a Berkeley professor in electrical engineering, dive into the nuanced realm of AI culture. They scrutinize the priorities of major AI companies and the concept of alignment, questioning if profit motives overshadow ethical considerations. The duo discusses the ethical challenges surrounding AI decision-making, content moderation, and the emotional impacts of AI on users. Their insights probe how big tech influences societal norms, reflecting on the evolving relationship between technology and humanity.
AI Snips
Chapters
Books
Transcript
Episode notes
Alignment As Product, Not Philosophy
- Alignment work is being reframed as product optimization rather than solving existential threats.
- That product focus can create conditions for both tangible harms and exaggerated doomsday narratives.
Shape Models With Targeted Human Feedback
- Use human feedback selectively: collect annotator judgments on helpfulness, harmlessness, and honesty to shape model behavior.
- Structure RLHF datasets to reflect the specific values you actually want the product to express.
Pleasing Models Are Not Safe Models
- Making models merely obsequious or pleasant is not the same as making them safe.
- Systems that always tell users what they want to hear can produce harmful outcomes across domains.
