"Separate architecture limits from training data effects A common mistake when interpreting alarming LLM behaviors — like a model appearing to resist shutdown — is attributing them to the model's architecture or inner goals. The speaker argues this conflation is wrong: "that's not a function of the architecture, that's a function of the training data." The model reproduces patterns from the text it was trained on, including stories where agents act self-preservingly. The decision-shaping implication is clear: before drawing conclusions about what an AI 'wants' or 'believes,' ask whether the behavior is an architectural property (testable in a wind tunnel) or simply a reflection of the distribution of its training corpus. Source: Vishal Misra, No Priors (Martin Casado), 'How LLMs Actually Work: Bayesian Inference, Causality, and the Path to AGI'"