Turing Award winner Judea Pearl discusses his career journey from electrical engineering and physics to pioneering Bayesian networks and causal reasoning in AI. He shares the eureka moment behind Bayesian networks — connecting graph theory to probability theory to handle uncertainty efficiently — and explains why probability alone proved insufficient, leading him to develop the ladder of causation (association, intervention, counterfactual explanation). Pearl places LLMs within this causal hierarchy, arguing they succeed by summarizing human-interpreted knowledge from the internet rather than raw data, but skeptically views current LLMs as insufficient for AGI without genuine causal understanding. He also reflects on early AI optimism in the 1960s-80s, the alpha-beta pruning optimality proof, and what curiosity-driven autonomy would mean for future AI systems.
9K Impressions1 Comment