Mereth Labs
← The research queue
Pure engineering

Scalability of personalisation

Bloom's 2 Sigma Problem

Bloom (1984): two performance distributions. The one-to-one tutored cohort sits two standard deviations above the conventional classroom.

In 1984, Benjamin Bloom reported a result that should have reorganised education and mostly didn’t. Students tutored one-to-one under mastery learning performed two standard deviations better than students in an ordinary classroom: the median tutored student outperformed 98% of the control group.

Δperformance≈2σ⟹P(tutored>medianclass)≈0.98 \Delta_{\text{performance}} \approx 2\sigma \quad\Longrightarrow\quad P(\text{tutored} > \text{median}_{\text{class}}) \approx 0.98

Why it stayed unsolved

The result was never in doubt. The problem was always delivery. One-to-one tutoring scales linearly with tutors: to give every student one, you need roughly one expert per student, at a cost no system has ever managed to pay. So education settled for the affordable approximation of thirty students to a teacher, taught in a single batch.

Why it is an engineering problem now

This is a problem of information distribution and compute scaling. The expertise already exists; what mattered was the marginal cost of delivering it, and that cost is now approaching zero. The constraint that kept the two-sigma result a curiosity instead of a blueprint has lifted, and closing the gap turns into an engineering job.