软件 2.0

Software 2.0

安德烈·卡帕西 Andrej Karpathy · Eureka Labs · 2017-11-11 · Karpathy Blog ↗

打开互动全文版（逐段中英对照 + 图/公式 + 论文问答）→

摘要 · Abstract

我有时看到人们将神经网络仅仅视为“机器学习工具箱中的另一个工具”。它们有优点和缺点，在某些地方有效，有时你可以用它们赢得 Kaggle 竞赛。不幸的是，这种解释完全只见树木不见森林。神经网络不仅仅是另一个分类器，它们代表了我们开发软件方式的根本性转变。它们是软件 2.0。我们熟悉的“经典栈”软件 1.0 是用 Python、C++ 等语言编写的。它由程序员编写的对计算机的显式指令组成。通过编写每一行代码，程序员在程序空间中确定一个具有某些期望行为的特定点。

[](https://karpathy.medium.com/?source=post_page---byline--a64152b37c35---------------------------------------) I sometimes see people refer to neural networks as just “another tool in your machine learning toolbox”. They have some pros and cons, they work here or there, and sometimes you can use them to win Kaggle competitions. Unfortunately, this interpretation completely misses the forest for the trees. Neural networks are not just another classifier, they represent the beginning of a fundamental shift in how we develop software. They are Software 2.0. The “classical stack” of Software 1.0 is what we’re all familiar with — it is written in languages such as Python, C++, etc. It consists of explicit instructions to the computer written by a programmer. By writing each line of code, the programmer identifies a specific point in program space with some desirable behavior.

核心贡献 · Key contributions

• 提出软件 2.0 作为一种新的编程范式，用神经网络替代显式代码。
  Proposes Software 2.0 as a new programming paradigm where neural networks replace explicit code.
• 表明对于许多现实问题，收集数据通常比编写显式程序更容易。
  Shows that collecting data is often easier than writing explicit programs for many real-world problems.
• 列出优势：计算同质性、恒定运行时间和内存、可移植性、敏捷性以及模块集成。
  Lists benefits: computational homogeneity, constant runtime and memory, portability, agility, and modular integration.
• 强调软件 2.0 在视觉、语音、翻译、游戏和数据库方面优于手工编码方案。
  Highlights that Software 2.0 outperforms hand-coded solutions in vision, speech, translation, games, and databases.
• 预测软件开发工具将转向以数据集为中心的 IDE 和包管理器。
  Predicts a shift in software development tools towards dataset-centric IDEs and package managers.
• 认为 AGI 将使用软件 2.0 原则构建。
  Argues that AGI will be built using Software 2.0 principles.

局限 · Limitations

• 软件 2.0 模型难以解释，在关键应用中导致信任问题。
  Software 2.0 models are hard to interpret, leading to trust issues in critical applications.
• 模型可能以不直观的方式失败，或默默吸收训练数据中的偏见。
  Models can fail in unintuitive ways or silently adopt biases from training data.
• 对抗样本揭示了传统软件中不存在的脆弱性。
  Adversarial examples reveal vulnerabilities that are not present in traditional software.
• 该范式需要大量标注数据集，创建成本高且耗时。
  The paradigm requires large labeled datasets, which are expensive and time-consuming to create.
• 与软件 1.0 IDE 相比，软件 2.0 的当前工具（如数据集管理）尚不成熟。
  Current tooling for Software 2.0 (e.g., dataset management) is immature compared to Software 1.0 IDEs.

论文章节 · Sections（共 7）

• 概述 Overview
• 正在进行的转变 Ongoing transition
• 软件 2.0 的优势 The benefits of Software 2.0
• 获取 Andrej Karpathy 的故事到你的收件箱 Get Andrej Karpathy’s stories in your inbox
• 软件 2.0 的局限性 The limitations of Software 2.0
• 在 2.0 栈中编程 Programming in the 2.0 stack
• 作者：Andrej Karpathy Written by Andrej Karpathy(https://karpathy.medium.com/?source=postpage---postauthorinfo--a64152b37c35---------------------------------------)

阅读逐段中英对照全文 →