Can some machine learning models handle unfamiliar data better than others? Léo Dreyfus-Schmidt discusses the limits of extrapolation and whether linear models might sometimes be more robust to data drift than decision tree-based approaches. The conversation explores: • Why decision trees struggle to extrapolate beyond what they’ve seen • Whether linear models can behave differently under data drift • Why no model can reliably extrapolate without making assumptions • The limits of predicting beyond the training distribution