"Spot the difference between association and causation in AI The speaker uses Judea Pearl's causal hierarchy to explain a hard ceiling on current deep learning: it operates at the level of association — finding correlations — but cannot perform intervention or counterfactual reasoning, which require a causal model. His concrete example: "when I throw this pen at you, your head is not doing a Bayesian calculation... you're doing a simulation" — something deep learning architectures cannot replicate. Anyone evaluating AI tools for high-stakes decisions (medical, legal, engineering) should ask whether the task requires only pattern recognition (where LLMs excel) or genuine causal reasoning (where they currently fail), and design human oversight accordingly. Source: Vishal Misra, No Priors (Martin Casado), 'How LLMs Actually Work: Bayesian Inference, Causality, and the Path to AGI'"