Read any webpage with AI. Annotate with Hypothesis in one click. 让 AI 为你读懂网页，一键标注到 Hypothesis。
~47 sec · 1440p · open any page → click the extension → pick a style → generate → check the ones to keep → publish → reopen the page to see annotations overlaid. 约 47 秒 · 1440p · 演示：打开任意网页 → 点扩展 → 选风格 → 生成候选 → 勾选发布 → 在原页面查看叠加标注
Some of the most important context is implicit, conditional, and historically contingent, and only exists as tribal knowledge inside teams.
Implicit context can't be fully encoded 隐性 context 不可被完全编码
The most important business context is implicit, conditional, and historically contingent — hard to fully capture or formalize. This pushes back on the fully-autonomous data-agent vision and underscores the irreplaceable role of humans in building context. 最重要的业务上下文往往是隐性的、有条件的、历史依赖的，难以被完全捕捉和编码。这挑战了完全自动化的数据代理愿景，强调了人类参与在上下文构建中的不可替代性。
View on Hypothesis → 在 Hypothesis 查看 →In this way the context layer can become a multi-dimensional corpus where code lives alongside natural language, capturing any context an agent might need.
Code + natural language as a dual-track corpus 代码 + 自然语言双轨知识库
The author proposes that the context layer should become a multi-dimensional corpus, blending code with natural language. The framing breaks past the traditional binary in data management and points to a new direction for genuinely intelligent data agents. 作者提出了一个创新性的概念：上下文层应成为多维度的知识库，将代码与自然语言融合。这一观点突破了传统数据管理的二元思维，为构建真正智能的数据代理提供了新思路。
View on Hypothesis → 在 Hypothesis 查看 →While model capabilities have improved dramatically for use cases like codegen and mathematical reasoning, they still lag behind on the data side (as evidenced through SQL benchmarks like Spider 2.0 and Bird Bench).
Strong at code ≠ strong at data (counterintuitive) 代码强 ≠ 数据强（反共识）
Despite dramatic gains on codegen and math, models still trail on data tasks. This challenges the "general capability uplift" assumption and suggests data reasoning may need its own approach. 尽管模型在代码生成和数学推理方面取得了显著进步，但在数据处理方面仍然落后。这挑战了模型能力全面提升的假设，暗示了数据推理可能需要特殊的处理方法。
View on Hypothesis → 在 Hypothesis 查看 →We are at an interesting point in time of market development, where the problem of a lack of context has become apparent, but we are still in the early innings of building solutions.
Problem identified · solutions still early 问题已识别 · 方案仍早期
The author's read of the market: the problem is identified, but solutions are still early. Implies a current gap between hype and capability — and meaningful innovation in the years ahead. 作者对市场发展阶段的分析提供了一个有洞察力的视角：问题已被识别，但解决方案仍处于早期阶段。这暗示了当前市场可能存在过度炒作与实际能力之间的差距，以及未来几年可能出现的实质性创新。
View on Hypothesis → 在 Hypothesis 查看 →📌 Install the Hypothesis browser extension and reopen the a16z article to see all 18 Chinese-language insights overlaid on the English original. 📌 装 Hypothesis 浏览器插件 后访问 a16z 原文，全部 18 条中文洞察会直接叠在英文正文上。
Annotating on Hypothesis amplifies high-quality reading, but the manual cost is high enough that most people give up. Hypothesisor reads the full page with GLM and, in the style you pick (non-consensus takes, gold sentences, critical reading, surprising facts…), drafts 5–40 quality candidates in one go. You check the ones you want and publish. Long pieces automatically get deeper coverage. 在 Hypothesis 上做标注是高质量阅读的放大器，但人工标注慢、容易放弃。Hypothesisor 用 GLM 读完整页正文，按你选的「风格」（非共识观点、金句、批判视角、AI 洞察…）一次生成 5-40 条有质量的候选标注，你勾选发布即可。长文会自动加深分析。
3-5 for shorts, 12-18 for long form, 18-28 for deep long form, 25-40 for chapter-length pieces. Coverage spans every section. 短文 3-5 条，长文 12-18 条，深度长文 18-28 条，超长篇章 25-40 条。覆盖各个章节。
Non-consensus / data / actionable / gold sentences / critique / surprising facts / any free-text custom prompt. 非共识观点 / 数据 / 行动 / 金句 / 批判性 / 令人惊讶的事实 / 任意自定义描述。
Page extraction, GLM call, and Hypothesis POST all happen in the extension. Tokens stay on your machine. 抓正文、调 GLM、POST Hypothesis 全在扩展里完成。Token 只存本地。
Only quotes that match the live DOM byte-for-byte get published — no AI "creative editing" of the source. 只有在当前页面正文中能逐字命中的引用才能发布，杜绝 AI "改编"原文。
Every annotation follows a "bold takeaway + explanation" structure, ≤120 chars. Renders natively in the Hypothesis UI. 每条标注统一「加粗点题 + 展开说明」结构，≤120 字，Hypothesis 原生渲染 Markdown。
Configure your Hypothesis Token and BigModel API Key once and you're set. No middleman. 配置一次 Hypothesis Token 和 BigModel API Key 就能用，没有中间商。
Clone the repo or download the ZIP: git clone https://github.com/fxp/Hypothesisor.git
克隆仓库或下载 ZIP：git clone https://github.com/fxp/Hypothesisor.git
Open chrome://extensions in Chrome → toggle "Developer mode" in the top right → click "Load unpacked" → choose the chrome-extension/ folder.
Chrome 打开 chrome://extensions → 右上角开「开发者模式」→ 点「加载已解压的扩展程序」→ 选 chrome-extension/ 目录。
Click the extension icon → ⚙ → fill in: 点扩展图标 → 齿轮 ⚙ → 填入：
On any page click the extension icon → pick mode/style → "Annotate this page" → check the candidates you want → "Publish selected". Install the Hypothesis browser extension and reopen the page to see the annotations overlaid. 在任意网页点扩展图标 → 选模式/风格 → 「生成标注」→ 勾选想发布的条目 → 「发布选中」。装 Hypothesis 浏览器插件后回到原页面即可看到叠加的标注。
| Article length 正文长度 | Tier 档位 | Annotations 标注条数 | Output tokens 输出 token |
|---|---|---|---|
| < 4k | Short短文 | 3–5 | 4096 |
| 4k – 12k | Medium中等 | 6–10 | 4096 |
| 12k – 30k | Long form长文 | 12–18 | 6144 |
| 30k – 60k | Deep long form深度长文 | 18–28 | 8192 |
| > 60k (truncated to 60k)（截至 60k） | Chapter-length超长篇章 | 25–40 | 8192 |
chrome-extension/
├── manifest.json # MV3 manifest
├── popup.{html,css,js} # main UI
├── options.{html,js} # settings
├── lib/agent.js # extract → GLM → validate → POST Hypothesis
├── icons/icon.svg
└── README.md
chrome-extension/
├── manifest.json # MV3 声明
├── popup.{html,css,js} # 主界面
├── options.{html,js} # Token 设置
├── lib/agent.js # 抽文本 → GLM → 验引用 → POST Hypothesis
├── icons/icon.svg
└── README.md
Page extraction: chrome.scripting.executeScript walks Text nodes under article / main / [role=main], byte-aligned with the Hypothesis client's anchoring layer. BigModel and Hypothesis APIs both called directly — no backend. Storage in chrome.storage.local.
抽文本：chrome.scripting.executeScript 读 article / main / [role=main] 的 innerText。BigModel 与 Hypothesis 两家 API 均直连，无后端。存储用 chrome.storage.local。