Manifold Dreamers and Doomers: Our AI future, with Richard Ngo – #109
25 snips
Apr 9, 2026 Richard Ngo, philosopher and independent AI researcher who worked on AGI safety at DeepMind and OpenAI. He discusses timelines and probabilities for advanced AI. He explores agent foundations, multi-agent views, and robustness/autonomy challenges. He debates communicating risk, lab cultures, governance and realistic takeover scenarios. Short, speculative, and provocative.
AI Snips
Chapters
Books
Transcript
Episode notes
Alignment Is Understanding Cognition, Not Band-Aids
- Ngo frames alignment research as the attempt to understand cognition's principles rather than just produce safe models.
- He positions agent foundations and mechanistic interpretability as central because they aim to reveal core cognitive mechanisms.
Multi-Agent Models Explain AI Internal Conflicts
- Treat AIs as multi-agent systems: internal sub-agents and sociological dynamics explain phenomena like the 'Waluigi effect'.
- Ngo suggests multi-agent concepts can bridge social science models to mechanistic interpretability of neural nets.
Protect Time To Think Unpublishable Thoughts
- Insulate exploratory research from institutional pressures to allow 'unpublishable' thoughts to form real breakthroughs.
- Ngo left labs to preserve intellectual freedom and pursue long-form philosophical ideas that may later yield publishable work.
