Puzzle of the day · Daily ML
Daily ML
One machine learning puzzle a day, the same for everyone. Compute a number, pick a metric, spot the mistake or name the paradox. A new puzzle appears at midnight UTC.
A report boasts 100% accuracy on the training data and gives no score for new data. What does it mean?
- The model is already perfect and ready
- The score is understated
- We know nothing about generalization
- The data are too simple
Show the answer
C. We know nothing about generalization
The model may have only memorized the samples. Only a score on data it has not seen shows whether it learned.
From the entry: Złudzenie trafności treningowej (PL)
Without JavaScript you see the puzzle of the day this page was published (2026-10-26); with JavaScript always today’s.
How it works
- No account and no ranking: you only compete with yesterday’s you.
- The result is coloured squares, one per try. They do not give the answer away, so you can paste them to friends.
- The streak is stored in this browser only and forgives one missed day.
- Every number is computed on the Atlas reference data or taken from the entry the puzzle links to.
The pool has 64 puzzles; none comes back until the whole pool has been played. Kinds of puzzle: 🔢 Compute · 📏 Metric or model · 🐞 Spot the mistake · ⚖️ Laws and paradoxes. Reference datasets.