The document content is in Chinese. Here’s the translation into English:

Andrej Karpathy’s speech highlights the AI industry’s most dramatic paradigm shift in 70 years, dividing software development into three stages: 1.0 (code era), 2.0 (neural network data programming), and 3.0 (large language models (LLMs) driving natural language interaction). He emphasizes that LLMs, as a new type of operating system, possess knowledge coverage, context memory, and cross-domain generalization capabilities, but have limitations such as hallucination and short-term memory deficits. The current LLM ecosystem shows dual tracks of closed-source (e.g., GPT-4) and open-source (e.g., Llama). Infrastructure needs transformation to support agent interaction, such as creating machine-readable documents and structured API interfaces. Karpathy advocates for “partially autonomous applications” to balance AI efficiency with human supervision, avoiding full automation traps. He warns that AI infrastructure requires transformation to meet agent needs, emphasizing the need for a decade of patience in AI development and cautioning against hype bubbles, while promoting the construction of a fundamental technology ecosystem.

Key Points:

  1. Three stages of software development: code era → data programming → natural language interaction-driven.
  2. LLMs as a new operating system with knowledge coverage, context memory, and cross-domain capabilities, but limitations like hallucination and short-term memory gaps.
  3. Infrastructure transformation required to support agent interaction, including machine-readable documents and structured API interfaces.
  4. “Partially autonomous applications” strategy to balance AI efficiency with human supervision, avoiding full automation traps.
  5. AI development needs a decade of patience, cautioning against hype bubbles, and promoting the construction of a fundamental technology ecosystem.

Translation

安德烈·卡帕西在演讲中指出，AI行业正经历70年来最剧烈的范式转变，软件发展分为三个阶段：1.0（代码时代）、2.0（神经网络数据编程）和3.0（大语言模型驱动的自然语言交互）。他强调大语言模型（LLM）作为新型操作系统，具备超越人类的知识储备和跨领域泛化能力，但存在幻觉、短期记忆局限等缺陷。当前LLM生态呈现闭源（如GPT-4）与开源（如Llama）双轨制，需通过“部分自治应用”构建人机协作模式，而非追求全自动化。卡帕西警告，AI基础设施需改造以适应智能体需求，例如创建机器可读文档、结构化API接口，并指出智能体发展需十年级的耐心，避免技术泡沫。

关键点：

  1. 软件发展三阶段：代码时代→数据编程→自然语言交互驱动。
  2. LLM作为新型操作系统，具备知识覆盖、上下文记忆和跨领域能力，但存在幻觉、短期记忆缺失等局限。
  3. 基础设施需改造以支持智能体交互，如创建机器可读文档、结构化API接口。
  4. “部分自治应用”策略平衡AI效率与人类监督，避免全自动化陷阱。
  5. AI发展需十年级耐心，警惕炒作泡沫，推动底层技术生态建设。

Reference:

https://www.youtube.com/watch?v=LCEmiRjPEtQ

<
Previous Post
Sam Altman on AGI, GPT-5, and what’s next
>
Next Post
Robotaxi vs Waymo and Cruise