A study buddy that refuses to give the answer¶
We built it in about an hour. Agreeing on what to build took most of the day.
This was at Claude Impact Lab in Oslo on 21 October. The idea was a study buddy for pupils, with one rule: it does not give the answer.
That sounds like a small rule. It is the opposite of what these models do by themselves. Ask one a maths question and you get the answer, worked out and explained, before you have thought at all. For homework that is close to useless. The pupil gets the answer and learns nothing.
So tenkemakker has a ladder of hints. The first is small, a nudge towards where to look. Ask again and you get a bit more. It moves up one step at a time, and it stops before the answer.
Then comes the part I like best. When the pupil has it, the roles are turned. The bot plays the one who does not understand, and the pupil has to teach it back. You do not know a thing until you can explain it, and a bot is a patient audience.
What I learned¶
Building it was the easy part. Once we were aligned on what it should do, it took an hour.
I am used to the opposite. In my work the idea is cheap and the building takes months, so we spend little time on the idea and a lot on the code. Here the cost has moved. When the building takes an hour, the time spent agreeing on a good idea is the work.
I found that uncomfortable, and I still do. The part I am good at is the part that just got cheap.