Summary
Panelists argued that scaling alone will not deliver reliable next-phase AI. They contrasted scientist AI for honest probabilistic guardrails, continual learning and test-time training, and world models that move from textual knowledge toward physical, social, and philosophical intelligence.
Key Topics
5 key topics from Nicholas Thompson, Eric Xing, Yoshua Bengio, Yuval Noah Harari and Yejin Choi at World Economic Forum Annual Meeting 2026. Thicker branches were mentioned more often in the talk.
Next Phase of Intelligence
Click a topic to open it
Key Takeaways
- Design AI objectives around honest prediction, not hidden goals.
- Set human-defined risk thresholds before deploying AI agents.
- Test continually learning systems after every significant model evolution.
- Train models to understand human norms before optimizing actions.
- Distinguish textual, physical, social, and philosophical intelligence needs.
“AI should really figure out how the world works for the sake of learning how the world works”
Share
Up Next
Next in agenda
Special Address by He Lifeng, Vice-Premier of China
He Lifeng, Berge Brende, Adnré Hoffmann
chinafree trade
Also on ai safety
The Day After AGI
Demis Hassabis, Dario Amodei, Zanny Minton Beddoes
agiai safety
Related
An Honest Conversation on AI and Humanity
Yuval Noah Harari, Irene Tracey
artificial intelligenceai ethics