GPT-3, with 175 billion parameters, demonstrated emergent few-shot learning, where it could perform new language tasks from only a few examples or simple instructions without explicit fine-tuning, showcasing surprising abilities like translation and question-answering not present in smaller models.