前特斯拉 AI 总监 Andre Carpathy 探讨软件从 1.0 到 3.0 的演变，强调神经网络和大语言模型如何改变编程方式。
Former Tesla AI director Andre Carpathy discusses the evolution of software from 1.0 to 3.0, highlighting how neural networks and LLMs are transforming programming.
要点 · TL;DR
软件 3.0 用自然语言提示作为程序，让编程人人可及。 Software 3.0 uses natural language prompts as programs, making coding accessible to everyone.
LLM 像人一样有记忆和认知缺陷，需要谨慎监督。 LLMs behave like humans with memory and cognitive flaws, requiring careful oversight.
构建部分自主产品，带快速人工验证循环，而非完全自主代理。 Build partial autonomy products with fast human verification loops, not full agents.
核心观点 · Key points
软件正在经历根本性变革，存在三种范式：1.0（代码）、2.0（神经网络）、3.0（提示词）。 Software is changing fundamentally with three paradigms: 1.0 (code), 2.0 (neural nets), 3.0 (prompts).
大语言模型类似于新的操作系统，上下文窗口相当于内存，提示词相当于程序。 LLMs are like new operating systems, with context windows as memory and prompts as programs.
大语言模型具有类人心理：百科全书式记忆，但也有幻觉、不均衡智能和顺行性遗忘。 LLMs have human-like psychology: encyclopedic memory but also hallucinations, jagged intelligence, and anterograde amnesia.
构建部分自主产品，配备图形界面以便人类快速验证，并设置自主性滑块。 Build partial autonomy products with a GUI for fast human verification and an autonomy slider.
氛围编码让任何人都能用自然语言编程，使软件开发人人可及。 Vibe coding lets anyone program in natural language, making software development accessible to all.
基础设施应适配 AI 智能体：使用 Markdown 文档，用 curl 命令替代点击操作。 Infrastructure should be adapted for AI agents: use markdown docs, replace clicks with curl commands.
反共识 · Contrarian takes
大语言模型不仅仅是公用事业；它们更像是具有复杂生态系统的操作系统。 LLMs are not just utilities; they are more like operating systems with complex ecosystems.
大语言模型的技术扩散方向是反向的：消费者先于政府和企业采用。 Technology diffusion for LLMs is reversed: consumers adopt before governments and corporations.
2025 年不是智能体元年；我们正处于智能体的十年，需要谨慎的人类监督。 2025 is not the year of agents; we are in the decade of agents, requiring careful human oversight.
让 AI 保持受控：小步快跑、快速验证优于大规模自主行动。 Keep AI on a leash: small incremental changes with fast verification loops are better than big autonomous actions.
为智能体构建基础设施比构建智能体本身更重要；与大语言模型相向而行。 Building for agents is more important than building agents themselves; meet LLMs halfway.
氛围编码是通往真正软件开发的入门途径，而非专业编码的替代品。 Vibe coding is a gateway drug to real software development, not a replacement for professional coding.
本期章节 · Chapters（共 28）
引言：AI 时代的软件Introduction: Software in the AI Era
软件 1.0、2.0 与 3.0Software 1.0, 2.0, and 3.0
示例：情感分类Example: Sentiment Classification
特斯拉自动驾驶：软件 2.0 吞噬全栈Tesla Autopilot: Software 2.0 Eating the Stack
聚焦 LLM 与新范式Focus on LLMs and the New Paradigm
LLM 作为公用设施LLMs as Utilities
LLM 作为晶圆厂与操作系统LLMs as Fabs and Operating Systems
更多类比：应用生态与分时系统More Analogies: App Ecosystem and Time-Sharing
LLM 颠覆技术扩散模式LLMs flip technology diffusion
LLM 如同 1960 年代的操作系统LLMs as operating systems circa 1960s
LLM 心理学：人的随机模拟Psychology of LLMs: stochastic simulations of people
超强记忆与认知缺陷Superhuman memory and cognitive deficits
安全局限：轻信与提示注入Security limitations: gullibility and prompt injection
机遇：Cursor 等部分自主应用Opportunities: partial autonomy apps like Cursor
LLM 应用特性Properties of LLM Apps
部分自主软件Partially Autonomous Software
与 AI 协作：生成与验证Cooperation with AI: Generation and Verification
AI 辅助编码的个人工作流Personal Workflow with AI-Assisted Coding
用课程给 AI 拴上缰绳Keeping AI on a Leash with Courses
特斯拉与 Waymo 的类比Analogy from Tesla Autopilot and Waymo
钢铁侠战衣类比Iron Man Suit Analogy
自然语言即编程与氛围编码Natural Language as Programming and Vibe Coding