ai-agent-book

GitHub/CommonMark does not treat the opening ** as an emphasis delimiter when it is preceded by normal Chinese text and immediately followed by Chinese punctuation such as “. In book/chapter1.md, replace:

因此它是**“构建”这一部分内部的一个观察视角**，而不是覆盖全书十章的目录。

with the reworded version from the fix:

因此，它是**构建部分内部的一个观察视角**，而不是覆盖全书十章的目录。

If the quotation marks must be preserved, use HTML instead:

因此它是<strong>“构建”这一部分内部的一个观察视角</strong>，而不是覆盖全书十章的目录。

Then reload the GitHub preview or published page to verify the phrase is bold.

Use an active, project-based approach instead of reading linearly. 1) Reimplement each chapter's experiments yourself rather than only reading the supplied code; a coding agent can help, but write and run the code. 2) Carefully observe experiment outputs and debug failures such as inaccessible APIs or unexpected results—these issues deepen understanding. 3) Apply the ideas to a real or personal AI-agent project. 4) Use the chapter thinking questions for self-testing. 5) Follow the cycle “first make the book thin, then thick, then thin again”: extract the core ideas, expand them with notes and experiments, then distill them into your own concise principles. The maintainers have also added these recommendations to the README FAQ; further reading questions can be discussed in the repository's Discussions area.

The description should be corrected. For each token, K_i = x_i @ W_k and V_i = x_i @ W_v depend only on that token's input and the projection weights; causal attention does not make K_i or V_i depend on previous tokens. Causal masking only determines which keys/values a token can attend to. When new tokens are appended, the existing prefix and its cached K/V remain unchanged, so the cache is still valid. The new token computes attention as z_i = softmax(Q_i · K_{0..i}^T / sqrt(d)) @ V_{0..i} using the cached prefix and its new entry. If an earlier token in the prefix changes, the affected cache entries must be recomputed.

The mismatch is caused by hardcoded chapter cards in the project root index.md. When manuscript titles or chapter order changed, the homepage 章节速览 cards were not updated, so Chapter 8 still displayed Agent 的持续进化 even though book/chapter8.md is titled 模型后训练.

Permanent fix: apply/merge the PR fix(site): generate homepage chapter index from manuscripts. Generate the homepage chapter cards during every MkDocs build from the current manuscript headings, using the configured navigation order and each edition's URL prefix and filename. Then rebuild and redeploy the site.

Temporary manual workaround: edit the hardcoded <span class="exp-title"> entries in index.md so they match the actual chapter headings and links. For Chapter 8, change the card title to 第 8 章 · 模型后训练. Also verify Chapters 5–9 because their titles/order appear shifted. Regenerate or update all localized homepage files (index.he.md, index.ko.md, index.ptbr.md, etc.) instead of fixing only the default homepage.

No. You do not need to understand every line before continuing. First understand the experiment's goal, architecture, and what each core component does. Read the chapter and pseudocode, then use a coding agent to implement the experiment from scratch until it produces the expected result. Study the implementation details only when you need fine-grained control over the behavior or want to learn the underlying language/framework. Focus on clearly describing requirements and design; let AI handle repetitive implementation code. The experiment READMEs and newly added pseudocode can help clarify the design.