The six sessions 📜 · from one black box to a choice you can defend
Two colours run through every figure and widget, so each culture always looks the same.
The data modelling culture
Assume a stochastic model generated the data, estimate its parameters, read nature's mechanism off the equation. Validated by goodness-of-fit: a yes or no.
The algorithmic culture
Treat the inside of the box as unknown, find a function that predicts y from x, and trust it only as far as its accuracy on unseen rows: a number.
Two ways to ask the data
Breiman's black box, the two cultures in his own words, the consulting projects that made him one of the 2 percent, and the hepatitis example where an unchecked model names the wrong variables. Build-along: both cultures on the same biopsies.
🧭 Session 2 · advancedAll models are wrong, some are useful
Where Box actually wrote it: the idea twice in 1976, the sentence as a 1979 section heading. What "useful" means as three checks you can run, and a regression that is 99.4 percent right with coefficients that point the wrong way.
🔎 Session 3 · advancedExplore first, confirm once
Tukey's 1962 case that data analysis is a science, exploratory work as detective work, and the discipline that keeps a confirmation honest: rows nobody touched until the end.
⚖️ Session 4 · bench nightThe two-cultures bench
A logistic regression and a random forest fitted in your browser on 569 real biopsies. Held-out accuracy and explanation size side by side, the base rate, a break button, and a peek switch that lies.
🗺️ Session 5 · advancedFifty years of data science
Donoho's 2017 history from Tukey to today, the common task framework he calls the secret sauce of machine learning, and the six divisions of greater data science, where both cultures share one division.
🤝 Session 6 · advancedWhich culture is this question?
Cox, Efron, Hoadley and Parzen in their own words, Breiman's rejoinder, and a way to choose a culture per problem that keeps their disagreement intact rather than pretending it was settled.
Your passport 🎫
Clear all three quiz questions in a session and it gets stamped. Six stamps means you can classify any model by culture, say what its validation does and does not license, keep a confirmation honest, and argue both sides of the 2001 exchange. Stamps live in this browser only.
What this is not, so you land in the right place 🧭
| If your question is | Go here instead |
|---|---|
| Why should I believe this model's test score, and what does a benchmark license? | ML Epistemology ↗ |
| Many models fit equally well and tell different stories. Which one do I trust? | ML Epistemology, leader session 2 ↗ |
| Frequentist or Bayesian, and how do I choose a prior? | Bayesian, leader session 5 ↗ |
| How does a random forest actually work? | Ensemble Methods, session 2 ↗ |
| How do I read the numbers my team brings me? | Data Literacy ↗ |
| How do I turn a business worry into a question data can answer? | Data Thinking ↗ |
| Where did the split between statistics and machine learning come from, what did each side actually claim, and how do I choose for a given problem? | you are in the right place |
Choose your path 🗺️
Three doors into the same conviction: a model is a claim, and the way it is checked decides what it lets you say.
What is honest here 🧾
The pages quote the papers themselves. Where a famous line could not be read in its original, the page says so and labels it reported.
| Claim | Status |
|---|---|
| Breiman 2001, and the comments by Cox, Efron, Hoadley and Parzen with Breiman's rejoinder | Read at source, Statistical Science 16(3), all 33 pages |
| Box: "Since all models are wrong" (1976) and "All models are wrong but some are useful" (1979) | Read at source, JASA 71(356) and Robustness in Statistics, Academic Press. The "essentially" version of 1987 is reported |
| Tukey 1962, The Future of Data Analysis | Read at source, all 67 pages. The 1977 book and the 1980 paper on exploratory and confirmatory analysis are reported |
| Donoho 2017, 50 Years of Data Science | Read at source, Journal of Computational and Graphical Statistics 26(4) |
| Every number on the bench | Really computed in your browser from the 569 real rows of UCI Breast Cancer Wisconsin (Diagnostic), and reproduced exactly by an independent Python run |
| "Culture X wins" | Not claimed. One dataset, nearly linearly separable; the pages teach the disagreement, not a verdict |
The knowledge map 🧠
The whole argument at a glance - hover a session to spotlight its concepts, click any node to jump in.