Learning a skill when a machine can already do it

This site is mostly about languages, and one part of it is about how learning works in general. This part is about the question that started arriving in 2023 and has not stopped: how do you learn to design, to program, to write, to analyse, when the thing you would have spent two years slowly getting good at is produced on demand, in seconds, to a standard that looks finished?

There is an answer, it is not the same answer for every skill, and almost everything written about it is either vendor copy or panic. These pages are neither. Each one sets out what actually changed about learning that skill, a staged path with an explicit rule at every stage about what the model may and may not do, the things you must still be able to do with it switched off — and the research, with what it measured and how much it can carry.

The one finding every page here rests on

In 2023 a research team gave nearly a thousand Turkish high-school students a GPT-4 tutor for their maths practice. While they had it, their practice performance rose 48%. Then it was taken away for the exam, and they scored 17% worse than the students who never had it at all. A second version of the same tutor, restricted to giving teacher-designed hints rather than answers, kept the gains and did no harm.

That is the shape of the whole problem, and it repeats in every discipline anyone has measured. Assistance improves what you produce while you have it. Whether it improves you depends entirely on what it was allowed to do — and the difference is invisible from the inside, because both feel like progress.

Two more numbers set the boundaries. Sixteen experienced developers, working on their own repositories, were measured taking 19% longer with AI tools while estimating afterwards that they had been made 20% faster: a forty-point gap between performance and self-assessment, among professionals. And 5,179 support agents gained 14% on average, 34% if they were novices and roughly nothing if they were experts — these tools transmit known practice downward far better than they create judgement.

Put together: the model is strongest exactly where you are weakest, and the point at which it stops being reliable is the point you are least equipped to notice. So the operating rule under every page below is the same one the hint-only tutor implemented — do the thinking first, unaided, and let the model near the work at the stage where you can already grade its answer.