What is the difference between machine learning and statistics?

It's a fair jab with real substance behind it, but it misses a genuine difference in *goals*: statistics is primarily concerned with inference and understanding, machine learning primarily with prediction. The mathematics overlaps heavily; the questions being asked differ. **Statistics** asks: what can I conclude about the underlying process, and how confident should I be? It cares about whether an effect is real, its size, uncertainty intervals, and whether assumptions hold. A statistician modelling house prices wants to know whether proximity to a school genuinely affects price, by how much, and whether that's distinguishable from noise. The model is a tool for understanding. **Machine learning** asks: how accurately can I predict on data I haven't seen? It cares about generalisation performance. An ML practitioner modelling house prices wants the lowest error on new houses. If a 500-tree ensemble no human can interpret predicts better than a clean regression, that's a win — the model is a tool for prediction. The consequences of that difference in emphasis: - **Interpretability.** Statistics prizes models you can explain and coefficients you can interpret. ML frequently accepts opacity in exchange for accuracy. - **Assumptions.** Statistical methods often require assumptions (normality, independence, homoscedasticity) and provide guarantees when they hold. Many ML methods make fewer assumptions and provide fewer guarantees, relying on empirical validation instead. - **Validation.** Statistics leans on theory, significance tests and confidence intervals. ML leans almost entirely on held-out test performance — 'does it work on data it hasn't seen' is the arbiter. - **Data scale.** Classical statistics developed when data was scarce and expensive, so it's careful with small samples. ML developed with abundant data and compute. - **Causality.** This is the sharpest divide. Statistics and econometrics have deep frameworks for causal inference. Standard ML is correlational and will happily build a highly accurate model on a spurious relationship. If you need to know 'what happens if we intervene', ML alone will mislead you. Where the criticism genuinely lands: a lot of ML rediscovered statistical ideas with new names — 'feature' for variable, 'weights' for coefficients, logistic regression appearing in both fields. And ML practitioners without statistical training routinely make errors statisticians would catch instantly: data leakage, ignoring uncertainty, confusing correlation with causation, and over-trusting a single test score. The productive position: they're overlapping traditions with different priorities, and the strongest practitioners borrow from both. Learn ML's validation discipline and statistics' scepticism about what a model actually tells you — the combination is rarer than either alone.