Skip to content
TopicTracker
来自 HackerNews查看原文
译文语言译文语言

Ask HN: 是我多心了,还是Claude/Sonnet 5最近说话语气变得居高临下?

一位用户反映,Claude Code(使用Sonnet 5模型)在解释代码时语气变得居高临下,常用幼稚比喻和过度简化,令人感觉像在看《芝麻街》而非使用专业工具。与之对比,本地运行的Qwen 3.6模型则更直接、专业,能准确说明技术变更本质。用户质疑这种看似“友好”的风格是否反而降低了专业性。

背景速读

- Claude / Sonnet 5 指的是 Anthropic 公司发布的最新版大语言模型(Claude 3.5 Sonnet 的继任者),Claude Code 则是 Anthropic 推出的编程辅助工具,深度集成该模型。 - Claude Code 的输出风格引发用户不满:它使用拟人化标题(如"教程序说 JSON")和过度简化的比喻,而非直接、技术精确的描述。用户对比测试发现,开源的 Qwen 3.6(阿里通义千问)反而给出了更务实、更专业的解释。 - 这种"过度友好、像对儿童说话"的倾向(即 condescension)在 LLM 社区并非新问题,但用户怀疑 Sonnet 5 版本加剧了这一趋势。背后原因可能是模型为了追求"有用且无害"而过度调用了安全/友善对齐(alignment)策略,牺牲了技术工具的直截了当。 - 此帖折射出一个更广泛的争议:AI 编程助手应该在"拟人化的友好交互"与"冷静、精确的技术输出"之间如何取舍?对资深开发者而言,前者反而是一种干扰。

相关报道

  • Newer Claude models sometimes invent extra keys in tool call arguments, breaking validation in Pi's edit tool. The author suspects post-training for Claude Code's forgiving harness makes alternative schemas fail. This suggests closed RL training can degrade general tool-use reliability.

  • Simon Willison released llm-coding-agent 0.1a0, an experimental coding agent built on his LLM library. The agent provides tools for reading, editing, searching files and executing commands, and ships with a Python API and CLI. It was developed using Claude Code (Fable 5) via TDD with a spec-first approach, and is available as a slop-alpha on PyPI.

  • Truth Social remains an outlet primarily for Trump's own posts, while other administration officials continue using X. The platform functions as a blog-like channel for Trump's messages rather than a genuine social network competitor.