feat: launch LLM Atlas research course
This commit is contained in:
wuyang
@@ -0,0 +1,298 @@
|
||||
---
|
||||
import BaseLayout from "@/layouts/BaseLayout.astro";
|
||||
import ArchitectureExplorer from "@/components/ArchitectureExplorer.astro";
|
||||
import DeepSeekLineage from "@/components/DeepSeekLineage.astro";
|
||||
import { chapters, statusLabel } from "@/data/chapters";
|
||||
|
||||
const routes: Record<string, string> = {
|
||||
roadmap: "/roadmap/",
|
||||
foundations: "/foundations/",
|
||||
};
|
||||
|
||||
const paths = [
|
||||
{
|
||||
id: "beginner",
|
||||
number: "01",
|
||||
label: "零基础",
|
||||
title: "从“下一个词”开始，不先背 Transformer 公式",
|
||||
text: "先建立 Token、概率、向量和训练目标的直觉，再通过可操作的注意力实验进入架构。",
|
||||
steps: ["01 语言模型从哪里来", "02 注意力与 Transformer", "03 表示、位置与残差", "04 Scaling Laws"],
|
||||
time: "8–12 小时",
|
||||
target: "能独立阅读主流 LLM 架构图",
|
||||
},
|
||||
{
|
||||
id: "k3",
|
||||
number: "02",
|
||||
label: "K3 反向拆解",
|
||||
title: "先看全貌，再沿组件回到每条技术祖先",
|
||||
text: "适合已经用过大模型、想迅速读懂 K3 报告的人。每个组件都能跳回其历史专题。",
|
||||
steps: ["K3 三维信息流", "KDA 与 MLA", "Stable LatentMoE", "1M Agentic RL 与系统"],
|
||||
time: "6–10 小时",
|
||||
target: "能逐节解释 K3 技术报告",
|
||||
},
|
||||
{
|
||||
id: "deepseek",
|
||||
number: "03",
|
||||
label: "DeepSeek 主线",
|
||||
title: "沿一家公司，看算法—系统协同如何一步步形成",
|
||||
text: "从 DeepSeek LLM 的 dense 基线，到 MoE、MLA、FP8、GRPO、R1、稀疏注意力与百万上下文。",
|
||||
steps: ["DeepSeekMoE", "V2 / MLA", "V3 / FP8 / DualPipe", "Math / GRPO / R1", "V3.2 / V4"],
|
||||
time: "7–11 小时",
|
||||
target: "理解 DeepSeek 论文间的因果关系",
|
||||
},
|
||||
{
|
||||
id: "systems",
|
||||
number: "04",
|
||||
label: "系统工程",
|
||||
title: "跟着一个 Token 穿过 GPU、网络、缓存与服务集群",
|
||||
text: "适合工程背景读者，从计算与内存账本出发理解并行、低精度、通信和推理服务。",
|
||||
steps: ["08 大规模训练", "09 数值与优化", "14 推理服务", "K3 §5 Infrastructure"],
|
||||
time: "10–16 小时",
|
||||
target: "看懂大模型系统报告与性能数字",
|
||||
},
|
||||
];
|
||||
---
|
||||
|
||||
<BaseLayout
|
||||
title="LLM Atlas"
|
||||
description="以 Kimi K3 技术报告为锚点，从第一性原理梳理大语言模型的技术发展、关键论文、DeepSeek 谱系与系统工程。"
|
||||
section="home"
|
||||
>
|
||||
<section class="hero">
|
||||
<div class="hero-inner">
|
||||
<div class="hero-copy">
|
||||
<p class="eyebrow"><span>OPEN COURSE / 2026</span> LARGE LANGUAGE MODELS</p>
|
||||
<h1>
|
||||
大模型技术全景
|
||||
<em>从“预测下一个词”到 Kimi K3 的 2.8T 开放前沿</em>
|
||||
</h1>
|
||||
<p class="hero-deck">
|
||||
这是一套按问题脉络组织的中文开放课程。我们从最小直觉出发，追踪每篇关键论文究竟解决了什么，
|
||||
再把注意力、MoE、规模化训练、多模态、推理强化学习与 Agent 系统重新汇入 Kimi K3。
|
||||
</p>
|
||||
<div class="hero-actions">
|
||||
<a class="button primary" href="/roadmap/">选择学习路径 <span aria-hidden="true">↓</span></a>
|
||||
<a class="button" href="/k3/">直接解剖 K3</a>
|
||||
<a class="button" href="/deepseek/">DeepSeek 专题</a>
|
||||
</div>
|
||||
</div>
|
||||
<aside class="hero-aside" aria-label="项目统计">
|
||||
<span>RESEARCH CUTOFF</span>
|
||||
<strong>28 · 07 · 2026</strong>
|
||||
<p>首版持续建设中，所有动态结论带日期</p>
|
||||
<div class="hero-stats">
|
||||
<div><b>16</b><span>核心专题</span></div>
|
||||
<div><b>151</b><span>K3 报告来源</span></div>
|
||||
<div><b>125</b><span>关键论文索引</span></div>
|
||||
<div><b>47p</b><span>K3 技术报告</span></div>
|
||||
</div>
|
||||
</aside>
|
||||
</div>
|
||||
<div class="scope-strip">
|
||||
<span>研究范围</span>
|
||||
<p>表示学习 · Transformer · Scaling · MoE · 长上下文 · 训练系统 · 后训练 · 推理 · Agent · 多模态 · 服务与评测</p>
|
||||
<a href="/progress/">查看实时进度 →</a>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="section" id="why">
|
||||
<div class="section-heading">
|
||||
<div>
|
||||
<p class="eyebrow"><span>00</span> WHY THIS ATLAS</p>
|
||||
<h2>不是论文清单，而是一条能走通的理解路径</h2>
|
||||
</div>
|
||||
<p class="section-lead">
|
||||
大模型知识最难的地方不是资料少，而是资料之间的桥断了：初学解释省略公式，论文省略历史，
|
||||
系统报告又默认你已经理解模型。LLM Atlas 专门补这些桥。
|
||||
</p>
|
||||
</div>
|
||||
|
||||
<div class="thesis">
|
||||
<span class="micro-label">课程主张</span>
|
||||
<p>
|
||||
真正理解一项技术，需要同时回答：它解决什么瓶颈、核心机制如何工作、证据是否支持、
|
||||
工程代价是什么，以及下一篇论文为什么自然会出现。
|
||||
</p>
|
||||
</div>
|
||||
|
||||
<div class="feature-grid">
|
||||
<article class="feature-card">
|
||||
<span class="card-number">01 / PROBLEM</span>
|
||||
<h3>从问题开始</h3>
|
||||
<p>先看到旧方法撞上的墙，再引入新名词。KDA、MLA、MoE 都不再是凭空掉下来的缩写。</p>
|
||||
</article>
|
||||
<article class="feature-card">
|
||||
<span class="card-number">02 / LAYERS</span>
|
||||
<h3>四层解释</h3>
|
||||
<p>同一概念同时提供直觉、机制、论文和工程四层入口，允许读者在适合自己的深度停留。</p>
|
||||
</article>
|
||||
<article class="feature-card">
|
||||
<span class="card-number">03 / VISUAL</span>
|
||||
<h3>图必须能帮助推理</h3>
|
||||
<p>优先重绘可缩放、可交互的架构图；每种颜色固定含义，简化之处明确标出。</p>
|
||||
</article>
|
||||
<article class="feature-card">
|
||||
<span class="card-number">04 / EVIDENCE</span>
|
||||
<h3>事实、解释、推断分开</h3>
|
||||
<p>关键结论回到论文和官方实现。榜单写明 harness、工具预算与截止日期，未知就明确说未知。</p>
|
||||
</article>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="section" id="paths">
|
||||
<div class="section-heading">
|
||||
<div>
|
||||
<p class="eyebrow"><span>01</span> LEARNING PATHS</p>
|
||||
<h2>你不必从第一页顺序读到最后</h2>
|
||||
</div>
|
||||
<p class="section-lead">四条路径共享同一张知识图。点击路径切换，章节之间的先修关系始终可追踪。</p>
|
||||
</div>
|
||||
|
||||
<div class="pathways" data-pathways>
|
||||
<div class="path-tabs" role="tablist" aria-label="学习路径">
|
||||
{paths.map((path, index) => (
|
||||
<button
|
||||
type="button"
|
||||
role="tab"
|
||||
aria-selected={index === 0 ? "true" : "false"}
|
||||
data-path-button={path.id}
|
||||
>
|
||||
<span>{path.number}</span>{path.label}
|
||||
</button>
|
||||
))}
|
||||
</div>
|
||||
<div>
|
||||
{paths.map((path, index) => (
|
||||
<div class="path-panel" data-path-panel={path.id} data-active={index === 0 ? "" : undefined}>
|
||||
<div class="path-copy">
|
||||
<span class="micro-label">{path.label}</span>
|
||||
<h3>{path.title}</h3>
|
||||
<p>{path.text}</p>
|
||||
<ol>
|
||||
{path.steps.map((step) => <li>{step}</li>)}
|
||||
</ol>
|
||||
</div>
|
||||
<aside class="path-meta">
|
||||
<div><span>ESTIMATED TIME</span><b>{path.time}</b></div>
|
||||
<div><span>LEARNING OUTCOME</span><b>{path.target}</b></div>
|
||||
<div><span>READING MODE</span><b>直觉 → 机制 → 论文 → 工程</b></div>
|
||||
</aside>
|
||||
</div>
|
||||
))}
|
||||
</div>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="section" id="chapters">
|
||||
<div class="section-heading">
|
||||
<div>
|
||||
<p class="eyebrow"><span>02</span> CURRICULUM MAP</p>
|
||||
<h2>16 个专题，拼成一张完整技术地图</h2>
|
||||
</div>
|
||||
<p class="section-lead">
|
||||
进度条表示内容成熟度，不代表相关领域的重要性。首版先打通全站骨架，再逐章扩写到论文与工程层。
|
||||
</p>
|
||||
</div>
|
||||
|
||||
<div class="chapter-grid">
|
||||
{chapters.map((chapter) => {
|
||||
const href = routes[chapter.slug] ?? `/roadmap/#chapter-${chapter.number}`;
|
||||
return (
|
||||
<a class="chapter-card" href={href}>
|
||||
<div class="chapter-top">
|
||||
<span class="chapter-number">{chapter.number}</span>
|
||||
<span class={`status ${chapter.status}`}>{statusLabel[chapter.status]}</span>
|
||||
</div>
|
||||
<h3>{chapter.title}</h3>
|
||||
<p>{chapter.question}</p>
|
||||
<div class="chapter-bottom">
|
||||
<div class="progress-track" aria-label={`完成 ${chapter.progress}%`}>
|
||||
<span style={`width:${chapter.progress}%`}></span>
|
||||
</div>
|
||||
<div class="chapter-meta">
|
||||
<span>{chapter.papers} 篇核心论文</span>
|
||||
<span>{chapter.progress}%</span>
|
||||
</div>
|
||||
</div>
|
||||
</a>
|
||||
);
|
||||
})}
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="section" id="k3-map">
|
||||
<div class="section-heading">
|
||||
<div>
|
||||
<p class="eyebrow"><span>03</span> KIMI K3 AS THE CONFLUENCE</p>
|
||||
<h2>K3 把三种信息流和一套庞大系统放进同一模型</h2>
|
||||
</div>
|
||||
<p class="section-lead">
|
||||
K3 报告最好的阅读钥匙，是把架构看成 token、depth、channel 三个方向的信息流；
|
||||
视觉输入与训练/服务基础设施则包住这三条轴。
|
||||
</p>
|
||||
</div>
|
||||
<ArchitectureExplorer />
|
||||
<a class="button primary" href="/k3/">进入完整 K3 技术报告导读 →</a>
|
||||
</section>
|
||||
|
||||
<section class="section" id="deepseek-spotlight">
|
||||
<div class="section-heading">
|
||||
<div>
|
||||
<p class="eyebrow"><span>04</span> DEEPSEEK SPOTLIGHT</p>
|
||||
<h2>DeepSeek：一条极适合学习“算法—系统协同”的论文主线</h2>
|
||||
</div>
|
||||
<p class="section-lead">
|
||||
从细粒度 MoE、MLA 和 FP8，到 GRPO 与 R1，DeepSeek 的每篇报告都在解决上一代留下的明确约束。
|
||||
这里会给予它比普通模型谱系更细的篇幅。
|
||||
</p>
|
||||
</div>
|
||||
<DeepSeekLineage />
|
||||
<a class="button primary" href="/deepseek/">阅读 DeepSeek 专题 →</a>
|
||||
</section>
|
||||
|
||||
<section class="section compact" id="method">
|
||||
<div class="section-heading">
|
||||
<div>
|
||||
<p class="eyebrow"><span>05</span> OPEN RESEARCH</p>
|
||||
<h2>课程本身也是一项开放研究工程</h2>
|
||||
</div>
|
||||
<p class="section-lead">
|
||||
网站源码、路线、进度和研究规范全部公开。Grok 用来扩展检索线索，最终结论由一手论文和官方实现核验。
|
||||
</p>
|
||||
</div>
|
||||
<div class="method-grid">
|
||||
<article class="method-card">
|
||||
<span class="card-number">P0</span>
|
||||
<h3>一手论文</h3>
|
||||
<p>arXiv、会议、期刊与作者正式技术报告，承载核心机制和实验数字。</p>
|
||||
</article>
|
||||
<article class="method-card">
|
||||
<span class="card-number">P1</span>
|
||||
<h3>官方实现</h3>
|
||||
<p>仓库、模型卡、训练与评测代码，用于确认论文描述如何落到真实系统。</p>
|
||||
</article>
|
||||
<article class="method-card">
|
||||
<span class="card-number">P2</span>
|
||||
<h3>官方说明</h3>
|
||||
<p>实验室博客、系统卡和产品文档；尚无论文的新模型会临时使用并标明状态。</p>
|
||||
</article>
|
||||
<article class="method-card">
|
||||
<span class="card-number">P3+</span>
|
||||
<h3>独立复现</h3>
|
||||
<p>第三方结果用于检验外部有效性；二手文章只发现线索，不承载关键事实。</p>
|
||||
</article>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<script>
|
||||
document.querySelectorAll<HTMLElement>("[data-pathways]").forEach((root) => {
|
||||
const buttons = [...root.querySelectorAll<HTMLButtonElement>("[data-path-button]")];
|
||||
const panels = [...root.querySelectorAll<HTMLElement>("[data-path-panel]")];
|
||||
const select = (id: string) => {
|
||||
buttons.forEach((button) => button.setAttribute("aria-selected", String(button.dataset.pathButton === id)));
|
||||
panels.forEach((panel) => panel.toggleAttribute("data-active", panel.dataset.pathPanel === id));
|
||||
};
|
||||
buttons.forEach((button) => button.addEventListener("click", () => select(button.dataset.pathButton ?? "beginner")));
|
||||
});
|
||||
</script>
|
||||
</BaseLayout>
|
||||
Reference in New Issue
Block a user