“A small dataset of approximately 30 examples was sufficient to teach early versions of ChatGPT about its own limitations, and this learning generalized effectively to capabilities not covered in the training data.”