本文旨在帮助你理解 AI 对工作、教育及生活的影响。作者：Ethan Mollick 教授。订阅即表示你同意 Substack 的服务条款，并知悉其信息收集通知与隐私政策。
Trying to understand the implications of AI for work, education, and life. By Prof. Ethan Mollick By subscribing, you agree to Substack's Terms of Use, and acknowledge its Information Collection Notice and Privacy Policy.
核心贡献 · Key contributions
指出智能体式系统是主要转变，将 AI 模型与工具结合以完成数小时工作。 Marks agentic systems as the main shift, combining AI models with tools to complete hours of work.
推荐 ChatGPT 和 Claude 作为实际工作的仅有两个实用选择，并给出具体模式与模型。 Recommends ChatGPT and Claude as the only two practical choices for real work, with specific modes and models.
强调权限设置和默认审批对于防止错误及提示注入攻击至关重要。 Highlights permission settings and approval defaults as essential to prevent mistakes and prompt injection attacks.
展示智能体在校对、电脑修复及操作电脑方面的能力，同时强调需人工审查。 Demonstrates agent capability in proofreading, computer fixes, and computer use, while stressing human review.
指出 Google 的 NotebookLM 和 Gemini Omni 是有用的特定工具，而 Copilot 和开放权重模型落后。 Identifies Google's NotebookLM and Gemini Omni as useful niche tools, while Copilot and open-weights lag behind.
建议支付 20 美元，委派一个真实任务，并像管理人一样与智能体迭代。 Advises paying 20 dollars, delegating one real task, and iterating with the agent like managing a person.
局限 · Limitations
20 美元档的智能体用量有限，更贵的套餐主要是购买更多工时而非更智能的 AI。 Agent usage at the $20 tier is limited, and costlier plans buy more hours rather than smarter AI.
提示注入和不受限的电脑操作仍是安全担忧，尽管模型抗性有所提升。 Prompt injection and unfettered computer use remain security concerns despite improved model resistance.
这些系统文档匮乏且命名混乱，使选择和导航变得不必要地复杂。 Systems are badly documented and confusingly named, making selection and navigation unnecessarily complicated.
人工监督至关重要，但在长期自主任务中可能成为瓶颈。 Human oversight is essential but may become a bottleneck for long-horizon autonomous tasks.
Google、Copilot 和开放权重模型在智能体式能力上有明显差距；开放权重需要专业知识才能有效使用。 Google, Copilot, and open-weight models have notable gaps in agentic abilities; open weights require expertise.
论文章节 · Sections（共 6）
One Useful Thing（一件有用的事）One Useful Thing(https://www.oneusefulthing.org/)