Andrej Karpathy：从 Vibe Coding 到 Agentic EngineeringAndrej Karpathy: From Vibe Coding to Agentic Engineering
The Takeaway: AI 正在从辅助工具转变为新的计算范式，我们必须重新思考编程和构建软件的方式。
Andrej Karpathy，这位 OpenAI 联合创始人、前 Tesla Autopilot 负责人，以其清晰解释复杂 AI 概念而闻名。他分享了自己从去年使用早期 agentic 工具到今年十二月体验的根本转变：最新模型能可靠地生成大块正确代码，他几乎不再需要修正。这让他开启了大量 side projects，并深刻意识到 LLM 已成为可编程的计算机。
Software 1.0 是显式规则，2.0 是训练权重，而 3.0 是通过提示和上下文窗口编程。Karpathy 用 OpenCLaw 安装和 MenuGen 应用举例：在新范式中，你不再编写复杂 shell 脚本，而是提供提示给 agent，让它智能处理环境差异。他强调，新机会在于以前不可能实现的事物，比如从文档自动创建知识库，而非仅加速现有流程。
他讨论了 jagged intelligence：模型在可验证领域（如代码、数学）表现出色，但在常识场景中仍显笨拙。"You can outsource your thinking, but you can't outsource your understanding." 人类仍需提供品味、判断和高层指导。未来将转向 agent-native 基础设施，一切为 agent 设计。
Karpathy 区分了 vibe coding（降低门槛）和 agentic engineering（在保持质量的同时加速专业软件开发）。
Andrej Karpathy，这位 OpenAI 联合创始人、前 Tesla Autopilot 负责人，以其清晰解释复杂 AI 概念而闻名。他分享了自己从去年使用早期 agentic 工具到今年十二月体验的根本转变：最新模型能可靠地生成大块正确代码，他几乎不再需要修正。这让他开启了大量 side projects，并深刻意识到 LLM 已成为可编程的计算机。
Software 1.0 是显式规则，2.0 是训练权重，而 3.0 是通过提示和上下文窗口编程。Karpathy 用 OpenCLaw 安装和 MenuGen 应用举例：在新范式中，你不再编写复杂 shell 脚本，而是提供提示给 agent，让它智能处理环境差异。他强调，新机会在于以前不可能实现的事物，比如从文档自动创建知识库，而非仅加速现有流程。
他讨论了 jagged intelligence：模型在可验证领域（如代码、数学）表现出色，但在常识场景中仍显笨拙。"You can outsource your thinking, but you can't outsource your understanding." 人类仍需提供品味、判断和高层指导。未来将转向 agent-native 基础设施，一切为 agent 设计。
Karpathy 区分了 vibe coding（降低门槛）和 agentic engineering（在保持质量的同时加速专业软件开发）。
The Takeaway: AI is shifting from a helpful tool to a new computing paradigm, forcing us to rethink how we program and build software.
Andrej Karpathy, OpenAI co-founder and former Tesla Autopilot lead, known for making complex AI concepts accessible, shared his stark transition from using early agentic tools last year to a fundamental change in December: latest models reliably produce large chunks of correct code with almost no corrections needed. This sparked a flurry of side projects and the realization that LLMs are now programmable computers.
Software 1.0 was explicit rules, 2.0 learned weights, and 3.0 is programming via prompts and context windows. Using examples like OpenCLaw installation and his MenuGen app, Karpathy illustrated that instead of complex shell scripts, you now provide prompts to agents that intelligently handle environment differences. New opportunities lie in things previously impossible, like automatically creating knowledge bases from documents, beyond just speeding up existing processes.
He discussed jagged intelligence: models excel in verifiable domains like code and math but remain clumsy in commonsense scenarios. "You can outsource your thinking, but you can't outsource your understanding." Humans must still provide taste, judgment, and high-level direction. The future moves toward agent-native infrastructure designed first for agents.
Karpathy distinguishes vibe coding (raising the floor) from agentic engineering (accelerating professional software while maintaining quality).
查看原文 →
Andrej Karpathy, OpenAI co-founder and former Tesla Autopilot lead, known for making complex AI concepts accessible, shared his stark transition from using early agentic tools last year to a fundamental change in December: latest models reliably produce large chunks of correct code with almost no corrections needed. This sparked a flurry of side projects and the realization that LLMs are now programmable computers.
Software 1.0 was explicit rules, 2.0 learned weights, and 3.0 is programming via prompts and context windows. Using examples like OpenCLaw installation and his MenuGen app, Karpathy illustrated that instead of complex shell scripts, you now provide prompts to agents that intelligently handle environment differences. New opportunities lie in things previously impossible, like automatically creating knowledge bases from documents, beyond just speeding up existing processes.
He discussed jagged intelligence: models excel in verifiable domains like code and math but remain clumsy in commonsense scenarios. "You can outsource your thinking, but you can't outsource your understanding." Humans must still provide taste, judgment, and high-level direction. The future moves toward agent-native infrastructure designed first for agents.
Karpathy distinguishes vibe coding (raising the floor) from agentic engineering (accelerating professional software while maintaining quality).