Prompt injection is a difficult problem that is hard to protect against. Despite delimiters being used to separate untrusted input, it can still be defeated, as demonstrated in the ChatGPT Prompt Engineering for Developers course. An attack on the model can trick it into thinking an instruction has already been completed, then tell it to do something else.
Helping a child who loves math to work through hard problems involves having them walk you through how they did the problem, without immediately pointing out their mistakes. This approach helps them develop their double checking skills, expanding their math universe and solidifying what they know and don’t know.
💡 By having children walk you through the problem step-by-step, it helps them to verbalize their process, redo all of the calculations, review the problem, and ultimately develop their “double checking” skills.