We're getting into semantics now, but that paper doesn't show learning. GPT can mimic learning capabilities through pattern recognition, but it works by feeding multiple past prompts in as one aggregate. If you show it examples, then go over the token limit (I believe it's 2048), it will "forget" its novel capabilities, because although the paper calls it few-shot learning, the text input is always zero-shot. All the past prompts are fed in again with every new message. The model stays static.