AI

Personalisation at scale: the open problem

One-to-one tutoring produces roughly a two-standard-deviation gain over conventional teaching. We've known that since 1984. The problem was never whether it works. It's how to give it to everyone.

Benjamin Bloom named it in 1984 and it has hung over the field ever since. Give a student one-to-one tutoring and their results jump by about two standard deviations over a conventional class. The average tutored student ends up ahead of 98% of the others. It’s one of the largest effects in all of education.

And it’s been almost useless — because you can’t give every learner a personal tutor. The result is real, and the economics are impossible. That gap is the two-sigma problem.

Why AI makes this live again

For the first time, the gap looks bridgeable. Not because a model can replace a great tutor — it can’t — but because a lot of what a good tutor does is structural: notice where you are, pitch the next step just beyond it, ask the question that makes you think, come back at the right moment. Some of that can now be delivered to millions at once.

The result was never in doubt. The delivery was. That’s what’s changing.

Where it’s real, and where it’s theatre

Plenty of “AI personalisation” is theatre — a chatbot bolted onto the same flat content, recommending the next video. That isn’t tutoring. It’s a recommendation engine wearing a lanyard.

The real version is harder and quieter. It adapts the difficulty and the timing, not just the playlist. It pushes people up Bloom’s taxonomy instead of drilling them at the bottom. It uses what it knows about a learner to stretch them, not to make things frictionlessly easy.

The honest state of it

We are not there yet. This is genuinely an open problem, and anyone claiming it’s solved is selling something. But it’s the most important one in learning technology, and for the first time in forty years the shape of an answer is visible. That’s what I spend my time on.