📄 中文摘要
提到“提示工程”,许多人首先想到的是如何向AI提问。然而,这种观点仅限于聊天用户的视角。对于开发者而言,提示不仅是提问,更是编程模型行为的过程。文章总结了在开发侧项目时学习到的几种提示技术,包括Zero-shot、Few-shot和Chain-of-Thought。Zero-shot是直接给出任务而不提供示例,适用于模型已知的任务;Few-shot则是提供几个示例以增强模型理解;Chain-of-Thought则是通过逐步推理来提高复杂任务的准确性。最后,ReAct被提出作为Chain-of-Thought的进化版本,强调思考、行动和观察的过程。
📄 English Summary
ChatGPT에 \"잘 물어보기\"는 아마추어다
The concept of 'prompt engineering' is often associated with asking AI questions effectively, but this perspective is limited to chat users. For developers, prompting is about programming the model's behavior. Various prompting techniques learned while developing side projects are summarized, including Zero-shot, Few-shot, and Chain-of-Thought. Zero-shot involves giving a task without examples, suitable for known tasks; Few-shot provides several examples to enhance understanding; Chain-of-Thought specifies step-by-step reasoning to improve accuracy in complex tasks. Finally, ReAct is introduced as an evolved version of Chain-of-Thought, emphasizing the processes of thinking, acting, and observing.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等