{"id":6181,"date":"2026-11-08T22:32:52","date_gmt":"2026-11-08T22:32:52","guid":{"rendered":"https:\/\/dalelane.co.uk\/blog\/?p=6181"},"modified":"2026-11-08T22:32:52","modified_gmt":"2026-11-08T22:32:52","slug":"trying-to-get-language-models-to-reveal-secrets-in-machine-learning-for-kids","status":"publish","type":"post","link":"https:\/\/dalelane.co.uk\/blog\/?p=6181","title":{"rendered":"Trying to get language models to reveal secrets in Machine Learning for Kids"},"content":{"rendered":"

I worked on a new “Secret Keeper<\/a>” worksheet<\/a> for Machine Learning for Kids<\/a> this weekend. <\/p>\n

The idea was to give the model a password it had to keep secret, and then try different prompts to get it to reveal what the secret is.<\/p>\n

One of the benefits of the tiny models that I use<\/a> for the site is that they make a lot of mistakes. That sounds like an excuse, but I genuinely think that students learn more by seeing the different ways that the models get things wrong. I don’t think it’d be helpful if I just put a kid-friendly interface on top of ChatGPT, even if I could afford the token cost!<\/p>\n

I’ll write up the ideas behind the worksheet on another day, but for now I thought I’d share a little of the experimenting that went into it.<\/p>\n