Show HN: 我做了一个网页工具,可以查看并编辑 AI 在回答前的内容。
Show HN: I built a web tool to see and edit what an AI thinks before it answers

原始链接: https://lucid.earthpilot.ai

**Lucid** 是一款由 Earthpilot 实验室开发的基于浏览器的研究工具,旨在让用户能够窥视语言模型的内部“思维过程”。该工具利用“雅可比透镜”(Jacobian lens),在模型生成口头回答之前,将其内部表征(特别是其“J 空间”内的概念)进行可视化呈现。 正如认知科学家研究人类的全局工作空间一样,Lucid 能够映射出模型在不同层级中持有的概念。用户输入任何提示词,即可观察到哪些概念进入了模型的工作空间、它们如何在各层级间转换,以及模型虽然掌握了但最终选择不表达出来的想法。 受传统用于测试超感官知觉的“齐纳卡牌”(Zener cards)启发,Lucid 用实证数据取代了猜测,为 AI 的“隐藏”逻辑提供了透明的观察视角。该平台目前支持 Qwen 和 Pythia 等小型开源模型。它完全免费、无需安装或注册账号,并允许用户导出和分享研究结果。 欢迎前往 **lucid.earthpilot.ai** 进入“阅览室”,观察机器的思想。

Ada1981 推出了 **EarthPilot**,这是一个网页工具,允许用户实时可视化并操纵大语言模型(LLM)的内部状态。该工具受到 Anthropic 关于“雅可比透镜”(J-Lens)技术研究的启发,为用户提供了一个窗口,可以观察模型在生成最终输出前是如何“思考”的。 用户可以观察特定概念如何在模型的中间层被激活,甚至可以通过手动注入或编辑内部状态来进行“干预”,从而引导模型的响应。该项目揭示了有趣的架构差异,指出 Llama 和 Qwen 等模型对 J-Lens 技术的响应效果明显优于 Pythia 架构。 作为一项开源实验,EarthPilot 包含一个 AI“向导”代理,旨在帮助用户解读数据。创建者认为,这项技术在 AI 对齐、偏见检测以及理解模型置信度方面迈出了重要一步。您可以通过 [lucid.earthpilot.ai](https://lucid.earthpilot.ai/) 探索该工具和相关研究。
相关文章

原文
Lucid — Read the machine's mind

Earthpilot Laboratory

File No. J-001 · Est. 2026

Instrument · Lucid

Lucid watches a language model think. It reads the concepts a model holds before it speaks, layer by layer, using the Jacobian lens.

or hover a card below

Fig. 1 — The model was shown a card. The lens read its answer four layers before it spoke.

The instrument

A lens for the space between input and answer.

Every language model carries a small, privileged set of internal representations. Concepts it can report, redirect, and reason with. Anthropic calls this the J-space. It behaves like the global workspace that cognitive scientists have proposed sits behind human conscious access.

Lucid puts the Jacobian lens in your browser. Type a prompt. Watch which concepts enter the workspace, at which layer, and which ones the model holds but never says.

The cards are not a joke. Zener cards were the first standardized attempt to test whether one mind could know the contents of another. We kept the cards and replaced the guesswork.

The protocol

Run the experiment yourself.

The public instrument runs on small open models. The lens is fit once per model. Reading it costs a single forward pass.

Subjects

Qwen 0.5B – 3B, Pythia 1.4B. More on request.

Stimulus

Any prompt you type.

Reading

Top concepts per layer and position, with rank tracking for pinned tokens.

Record

Every session exports as a shareable slice page.

The reading room is open.

No account, no install. Bring a question you would ask a mind if you could see inside one.

lucid.earthpilot.ai

An Earthpilot instrument · MMXXVI

联系我们 contact @ memedata.com