从「氛围编程」到智能体工程

From vibe coding to agentic engineering

安德烈·卡帕西 Andrej Karpathy · Training Data · 2026-04-29 · 约 30 分钟 · 原视频 ↗

本期速览 · Overview

当智能体来写代码、而你来「工程化」这个智能体，会有什么不同。

What changes when agents write the code and you engineer the agent.

软件 3.0 将编程从写代码转向通过上下文窗口提示大模型。
Software 3.0 shifts programming from writing code to prompting LLMs via context windows.
大模型在可验证任务（如代码）上表现出色，但在简单常识上失败。
LLMs excel in verifiable tasks like code but fail on simple common sense.
可以外包思考，但理解仍是人类瓶颈。
Outsourcing thinking is possible, but understanding remains a human bottleneck.

软件 3.0 是通过提示进行编程；上下文窗口是你对 LLM 解释器的杠杆。
Software 3.0 is programming by prompting; the context window is your lever over the LLM interpreter.
LLM 是锯齿状的：它们在数学和代码等可验证领域表现出色，但在步行去洗车等简单常识上失败。
LLMs are jagged: they excel in verifiable domains like math and code, but fail on simple common sense like walking to a car wash.
氛围编码提升了所有人的下限；智能体工程为专业人士保持质量门槛。
Vibe coding raises the floor for everyone; agentic engineering preserves quality bar for professionals.
我们召唤的是幽灵，而非构建动物——这些是由数据和奖励塑造的统计模拟，而非进化产物。
We are summoning ghosts, not building animals—these are statistical simulations shaped by data and rewards, not evolution.
你可以外包思考，但无法外包理解——这仍然是人类的瓶颈。
You can outsource your thinking, but you can't outsource your understanding—that remains the human bottleneck.
一切最终都可自动化；可验证性决定了当前自动化的难易程度。
Everything is automatable eventually; verifiability determines what is easy or hard to automate now.

软件 3.0 意味着编程是复制粘贴文本给智能体，而不是写代码。
Software 3.0 means programming is copy-pasting text for agents, not writing code.
许多现有应用如 MenuGen 是多余的；神经网络可直接处理输入到输出，无需中间件。
Many existing apps like MenuGen are spurious; neural networks can directly process input to output without middleware.
大语言模型是幽灵而非动物；它们缺乏内在动机，纯粹由数据和奖励塑造。
LLMs are ghosts, not animals; they lack intrinsic motivation and are shaped purely by data and rewards.
招聘 AI 原生工程师应通过大型项目和对抗性测试，而非解谜题。
Hiring for AI-native engineers should involve large projects and adversarial testing, not puzzles.
你可以外包思考，但不能外包理解；理解仍然是人类的瓶颈。
You can outsource thinking but not understanding; understanding remains a human bottleneck.
大语言模型的能力参差不齐部分源于实验室对高价值领域的侧重，而非根本限制。
The jaggedness of LLMs is partly due to labs' focus on valuable domains, not fundamental limits.