Causal reasoning is what intelligence needs to be safe and super.
safe A system is safe when the probability of harm associated with its existence and use is reasonably low. Harm is a counterfactual notion: an undesired outcome that would not have occurred without the system. Safety therefore cannot be specified or checked without causal models.
super Causal reasoning, as formalized by Judea Pearl, is what lets a system work out the result of an intervention and reason about scenarios that differ in their conditions but share the same structural mechanisms. A system that does this worse than humans is not superintelligent.
Writing
- Reading an LLM's Mind with Causal Effectssafe
- Grow your agent wiselysuper
- Personalized Decision Makingsafe
- Counterfactual Fairness: take as much as you cansafe
- The Causal Mindset Frameworksuper
- How to produce Intervenable features inside LLMssafe
