OpenAI 推出了 Codex,其首个成熟的用于编码的 AI 代理

  • OpenAI has introduced Codex, an agentic coding tool in research preview.

    • It allows experienced developers to delegate tasks to an AI agent that generates code and shows its work.
  • Codex is a unique interface in the ChatGPT web app sidebar. Users can click "code" to generate code or "ask" to get answers and advice.
  • Each task is performed in a distinct container with the user's codebase to reflect the development environment.
  • Developers can include an "AGENTS.md" file to provide custom instructions for the AI agent.
  • Codex is built on codex-1, a fine-tuned variation of OpenAI's o3 reasoning model trained on coding tasks.
  • OpenAI's announcement post about Codex addresses common objections against AI coding agents.

    • Fine-tuning aims to address concerns like non-standard code, opacity, and insecurity.
    • Codex shows its thinking and work during tasks, but users still need to review and validate the code.
  • Codex is available in research preview and is rolling out to ChatGPT Pro, Enterprise, and Team users. Plus and Edu support will come later. Users have generous access for free in the coming weeks with potential rate limits and a new pricing scheme later.
阅读 18
0 条评论