[{"data":1,"prerenderedAt":595},["ShallowReactive",2],{"content-\u002Fposts\u002Fintroduction-to-large-language-model":3,"blog-posts":321},{"id":4,"title":5,"body":6,"categories":309,"date":310,"description":58,"draft":311,"extension":312,"meta":313,"navigation":206,"path":314,"seo":315,"stem":316,"tags":317,"__hash__":320},"posts\u002Fposts\u002Fintroduction-to-large-language-model.md","大语言模型入门基础",{"type":7,"value":8,"toc":302},"minimark",[9,13,18,21,24,27,30,33,37,40,43,46,49,52,136,139,142,278,282,285,288,292,295,298],[10,11,12],"p",{},"本文介绍大语言模型(LLM)的基础内容。",[14,15,17],"h2",{"id":16},"insight","Insight",[10,19,20],{},"对于推理时的 LLMs 来说，它只做一件事：预测下一个 token。这听起来是一件非常简单的一件事。为什么 LLMs 却在文本模态上表现的如此全能呢？",[10,22,23],{},"力大砖飞？道法自然？或许都对。LLMs 在海量数据上做了 pretrain。截止到 2026 年 4 月份，国内开源的几家模型公司在预训练阶段训练的 token 数量大约都在 30T-35TB 之间。模型每个参数预训练消耗的 token 都远大于 20 了。并且我相信这个数据使用量会在 DeepSeek V4 分享他们新的稀疏方法后，继续增长。另外，qwen 小模型的实践告诉了我们：预训练时，随着数据量的增多，即使是小模型，其能力也在稳步提升。",[14,25,26],{"id":26},"数据飞轮",[10,28,29],{},"在模型的长上下文突破(RoPE、YaRN), 基座模型 Agentic 能力提升， LLM Infra(vLLM、SGLang etc.) Day0 级跟进支持等等突破之后，一批全新的、过去没有的长上下文 Agent 任务的数据在飞速产生。这个时候谁拥有更多数据（并且能聪明地运用），那么他们下一代的模型就会更强，这是个正反馈的过程，也被称为数据飞轮。",[10,31,32],{},"数据飞轮正在拉大不同模型厂商之间的差距。",[14,34,36],{"id":35},"看到新的-token-产生的时候发生了什么","看到新的 Token 产生的时候，发生了什么？",[10,38,39],{},"使用 LLM 时，输入的是人类的文本，输出的是人类的文本，这个过程中模型到底做了什么呢？",[10,41,42],{},"在看到文本的前一步，token 是经过采样后得到的 next token ID，根据 ID 一做 detokenizer 就得到看见的文本了。",[10,44,45],{},"得到 token id 的前一步，是经过 Softmax 计算出的一个概率矩阵，根据不同的解码策略（贪婪解码：取概率最大的； 投机解码：需要 MTP 模块； 或者按概率采样等等），选择 next token id。再上一步就是 logits。我们平常说的 Temperature（对 logits 矩阵做除法）、Top-K（只在概率最大的 K 个 Token 中选）、Top-P（只在累积概率 P 以内的 Token 中选）都是在这步做的。",[10,47,48],{},"logits 输出前是 LM Head（Linear 层），它将 hidden Linear 的特征数扩大到词表的大小。在此之前是文本经过 Embedding 向量化和 PE（Positional Encode）后进入若干个 Transformer Block 后又过了一遍 RMSNorm。",[10,50,51],{},"Transformer Block 很简单。就是 Attn 机制加上前馈网络（FFN）罢了。",[53,54,59],"pre",{"className":55,"code":56,"language":57,"meta":58,"style":58},"language-python shiki shiki-themes github-light github-dark","# Input: x\n# Transformer Block\ndef forward(self, x):\n    x = x + Attn(RMSNorm(x))  # 简单的残差\n    x = x + FFN(RMSNorm(x))\n    return x\n","python","",[60,61,62,71,77,92,113,127],"code",{"__ignoreMap":58},[63,64,67],"span",{"class":65,"line":66},"line",1,[63,68,70],{"class":69},"sJ8bj","# Input: x\n",[63,72,74],{"class":65,"line":73},2,[63,75,76],{"class":69},"# Transformer Block\n",[63,78,80,84,88],{"class":65,"line":79},3,[63,81,83],{"class":82},"szBVR","def",[63,85,87],{"class":86},"sScJk"," forward",[63,89,91],{"class":90},"sVt8B","(self, x):\n",[63,93,95,98,101,104,107,110],{"class":65,"line":94},4,[63,96,97],{"class":90},"    x ",[63,99,100],{"class":82},"=",[63,102,103],{"class":90}," x ",[63,105,106],{"class":82},"+",[63,108,109],{"class":90}," Attn(RMSNorm(x))  ",[63,111,112],{"class":69},"# 简单的残差\n",[63,114,116,118,120,122,124],{"class":65,"line":115},5,[63,117,97],{"class":90},[63,119,100],{"class":82},[63,121,103],{"class":90},[63,123,106],{"class":82},[63,125,126],{"class":90}," FFN(RMSNorm(x))\n",[63,128,130,133],{"class":65,"line":129},6,[63,131,132],{"class":82},"    return",[63,134,135],{"class":90}," x\n",[10,137,138],{},"Attn 机制的话，现在有 DSA、MLA 等等。不过未来的趋势还是掌管稀疏的神——DeepSeek 的新 Attention 的时代，这成本降的太夸张了。",[10,140,141],{},"干了什么？混合注意力（CSA：压缩稀疏注意力， HCA：高度压缩注意力）。不仅仅是上一代注意力的稀疏性，还直接对 Attn 进行压缩。表明了一个潜在的观点，对于长上下文（百万级），即使利用其稀疏性能够将 Attn 的算法复杂度降到 O(L)级。但是面对长上下文存储的 KV Cache 还是无能为力。V4 直接对 Attn 进行压缩，减少 KV Cache 数量。（Tips: KV Cache 可以近似看成 Context 的物理实体）",[53,143,145],{"className":55,"code":144,"language":57,"meta":58,"style":58},"def forward(self, x):\n    # 计算多头注意力 -> xq, xk, xv\n    xq, xk = RMSNorm(xq), RMSNorm(xk) # qk-norm\n    \n    xq, xk = apply_RoPE(xq, xk, cos, sin) # 加入相对位置编码\n    #... 存 KV Cache。  \n    # KV Cache 一般存显存。但是现在有存在SSD上的，比如DeepSeek，使得缓存命中成本惊人的低，是未来降低推理成本的一个方向\n\n    output = F.scaled_dot_product_attention(xq, xk, xk, is_casual=True) # FlashAttention\n    output = output.transpose(1, 2).reshape(bsz, seq_lean, -1)\n\n    return output, past_kv\n",[60,146,147,155,160,173,178,190,195,201,208,235,265,270],{"__ignoreMap":58},[63,148,149,151,153],{"class":65,"line":66},[63,150,83],{"class":82},[63,152,87],{"class":86},[63,154,91],{"class":90},[63,156,157],{"class":65,"line":73},[63,158,159],{"class":69},"    # 计算多头注意力 -> xq, xk, xv\n",[63,161,162,165,167,170],{"class":65,"line":79},[63,163,164],{"class":90},"    xq, xk ",[63,166,100],{"class":82},[63,168,169],{"class":90}," RMSNorm(xq), RMSNorm(xk) ",[63,171,172],{"class":69},"# qk-norm\n",[63,174,175],{"class":65,"line":94},[63,176,177],{"class":90},"    \n",[63,179,180,182,184,187],{"class":65,"line":115},[63,181,164],{"class":90},[63,183,100],{"class":82},[63,185,186],{"class":90}," apply_RoPE(xq, xk, cos, sin) ",[63,188,189],{"class":69},"# 加入相对位置编码\n",[63,191,192],{"class":65,"line":129},[63,193,194],{"class":69},"    #... 存 KV Cache。  \n",[63,196,198],{"class":65,"line":197},7,[63,199,200],{"class":69},"    # KV Cache 一般存显存。但是现在有存在SSD上的，比如DeepSeek，使得缓存命中成本惊人的低，是未来降低推理成本的一个方向\n",[63,202,204],{"class":65,"line":203},8,[63,205,207],{"emptyLinePlaceholder":206},true,"\n",[63,209,211,214,216,219,223,225,229,232],{"class":65,"line":210},9,[63,212,213],{"class":90},"    output ",[63,215,100],{"class":82},[63,217,218],{"class":90}," F.scaled_dot_product_attention(xq, xk, xk, ",[63,220,222],{"class":221},"s4XuR","is_casual",[63,224,100],{"class":82},[63,226,228],{"class":227},"sj4cs","True",[63,230,231],{"class":90},") ",[63,233,234],{"class":69},"# FlashAttention\n",[63,236,238,240,242,245,248,251,254,257,260,262],{"class":65,"line":237},10,[63,239,213],{"class":90},[63,241,100],{"class":82},[63,243,244],{"class":90}," output.transpose(",[63,246,247],{"class":227},"1",[63,249,250],{"class":90},", ",[63,252,253],{"class":227},"2",[63,255,256],{"class":90},").reshape(bsz, seq_lean, ",[63,258,259],{"class":82},"-",[63,261,247],{"class":227},[63,263,264],{"class":90},")\n",[63,266,268],{"class":65,"line":267},11,[63,269,207],{"emptyLinePlaceholder":206},[63,271,273,275],{"class":65,"line":272},12,[63,274,132],{"class":82},[63,276,277],{"class":90}," output, past_kv\n",[14,279,281],{"id":280},"optimizer","Optimizer",[10,283,284],{},"除了 Embeding 模型和 LM Head Linear 层以及其余 Linear 层、Bias 用 AdamW 优化器，其余矩阵用 Muon(dim>=2)。Muon 能够保证矩阵正交性，可以加速收敛，Loss 下降更稳定。但是 Muon 较新，一些后端实现的性能不好（比如 PyTorch 2.12 的 mps 后端， 替换 AdamW 之后训练速度变慢了）。",[10,286,287],{},"不过使用 Muon 是趋势，相比 AdamW 能够减少一半 optim 的参数使用，训练也更稳定。",[14,289,291],{"id":290},"mtp-多-token-预测","MTP 多 Token 预测",[10,293,294],{},"一次推理输出多个 Token。",[10,296,297],{},"如何实现？加上一个 MTP 模块（可以当只有一块注意力块的 LLM）：输入是模型预测出的 t+1，经过类似的步骤但是 Transformer Block 只有一块（减少计算量以及显存开销），得到 t+2。嵌套调用后可以一次推理得到多个 token。和主模型的 logits 一对比后，结果一样的就保留，不一样的就重新预测。因为序列相近，语义大概率类似，因此成功率很高，一般可以实现 2-3 倍的加速比。",[299,300,301],"style",{},"html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html pre.shiki code .s4XuR, html code.shiki .s4XuR{--shiki-default:#E36209;--shiki-dark:#FFAB70}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}",{"title":58,"searchDepth":73,"depth":79,"links":303},[304,305,306,307,308],{"id":16,"depth":73,"text":17},{"id":26,"depth":73,"text":26},{"id":35,"depth":73,"text":36},{"id":280,"depth":73,"text":281},{"id":290,"depth":73,"text":291},"技术","2026-06-16 22:36:11",false,"md",{},"\u002Fposts\u002Fintroduction-to-large-language-model",{"title":5,"description":58},"posts\u002Fintroduction-to-large-language-model",[318,319],"Transformer","LLMs","djNRTWkC1BNsIbTT8ceJRFrVVT0vy5eELFswj6opXns",[322,331,337,344,349,358,367,373,377,386,395,401,407,415,419,423,433,440,446,450,454,462,469,477,482,488,493,495,500,507,514,521,527,534,542,548,554,560,566,571,580,588],{"path":323,"title":324,"description":325,"date":326,"categories":327,"tags":328,"draft":206},"\u002Fposts\u002Ftransposition-borrowed-chords","平移借用：大调里立刻能用的和弦","把大调整体平移，借用各调的顺阶和弦；有符号平移量 ±2、±3、±5、±7 给出七个近关系调，合计 25 个现成三和弦与 31 个现成七和弦，全部 12 个平移量给 36 与 48。","2026-11-08 16:20:00","音乐",[329,330],"乐理","和弦",{"path":332,"title":333,"description":334,"date":335,"categories":327,"tags":336,"draft":206},"\u002Fposts\u002Fdiatonic-chords-major-system","大调体系下的三和弦和七和弦","以同主音 C 为基准，逐个推出七套音阶的顺阶三和弦与七和弦，标注每个和弦的出处，去重后得到 21 个三和弦与 28 个七和弦。","2026-11-08 15:20:00",[329,330],{"path":338,"title":339,"description":58,"date":340,"categories":341,"tags":342,"draft":206},"\u002Fposts\u002F作息感慨","作息感慨","2026-11-04 07:15:08","生活杂谈",[343],"日常",{"path":345,"title":346,"description":58,"date":347,"categories":58,"tags":348,"draft":206},"\u002Fposts\u002Fgp-tab-测试","gp-tab-测试","2026-10-12 10:51:22",null,{"path":350,"title":351,"description":352,"date":353,"categories":309,"tags":354,"draft":311},"\u002Fposts\u002Fnetwork-clean","怎么让网络变得干净","一个干净的网络，让人省心的使用服务。","2026-10-11 19:40:02",[355,356,357],"VPN","网络","AI",{"path":359,"title":360,"description":361,"date":362,"categories":341,"tags":363,"draft":311},"\u002Fposts\u002Fthe-age-of-hesitation","[搬运] 不敢投入的时代","为什么我们现在越来越不敢投入了。很多时候我们以为的选择，比如你有两个工作offer，现在选择好像是我到底应该考公还是考编。明明这两个选项都不是真正的选项，因为你还没有开始，也不知道最后能不能成。这个时候我们就已经开始犹豫了。","2026-10-10 11:44:12",[364,365,366],"翻译","搬运","心理学",{"path":368,"title":369,"description":370,"date":371,"categories":341,"tags":372,"draft":206},"\u002Fposts\u002Fwhy-happiness-look-like-same","[搬运] 为什么大家的幸福都长一个样？","为什么大家的幸福都长一个样？一种认知被广泛传播后似乎变得不断扁平，这种感觉很奇怪。","2026-10-09 13:57:15",[364,365,366],{"path":374,"title":375,"description":58,"date":376,"categories":341,"tags":348,"draft":206},"\u002Fposts\u002Feverything-is-paid-twice","任何事情都要被消费两次","2026-10-06 12:30:06",{"path":378,"title":379,"description":380,"date":381,"categories":382,"tags":383,"draft":311},"\u002Fposts\u002Fllm","为什么模型上线后，体感像是“降智”了？","权重与 KV cache 的量化都能降低 serving 成本，也可能改变输出；但体感变弱不足以证明服务后来切换了精度。","2026-09-29 21:44:08","技术杂谈",[319,384,385],"Serving","量化",{"path":387,"title":388,"description":389,"date":390,"categories":309,"tags":391,"draft":311},"\u002Fposts\u002Fspinning-up-policy-optimization","[译] 深度强化学习入门（三）：策略优化入门","OpenAI Spinning Up 教程第三部分的全文翻译：推导最简单的策略梯度、log-derivative trick、EGLP 引理、reward-to-go 与基线，配 PyTorch 示例代码。","2026-09-28 17:49:25",[392,393,394,364],"强化学习","RL","Spinning Up",{"path":396,"title":397,"description":398,"date":399,"categories":309,"tags":400,"draft":311},"\u002Fposts\u002Fspinning-up-rl-intro-2","[译] 深度强化学习入门（二）：强化学习算法的分类","OpenAI Spinning Up 第二部分的全文翻译：无模型与基于模型的 RL、策略优化与 Q-learning 的取舍，以及规划、专家迭代和经验增强等方法。","2026-09-28 10:20:21",[392,393,394,364],{"path":402,"title":403,"description":404,"date":405,"categories":309,"tags":406,"draft":311},"\u002Fposts\u002Fspinning-up-rl-intro","[译] 深度强化学习入门（一）：RL 的关键概念","OpenAI Spinning Up 教程第一部分的全文翻译：智能体与环境、状态与观测、动作空间、策略、轨迹、奖励与回报、价值函数、贝尔曼方程、优势函数，附 MDP 形式化。","2026-09-28 09:29:40",[392,393,394,364],{"path":408,"title":409,"description":410,"date":411,"categories":341,"tags":412,"draft":206},"\u002Fposts\u002Fwhen-machines-come-alive","冰箱之后，生活还会怎样改变？","从厨房里的果蝇想到冰箱、移动互联网与 AI 的多轮调用：AGI 已经到来，只是昂贵的智能会怎样改变生活？","2026-09-25 06:36:22",[357,413,414],"随笔","科技",{"path":416,"title":417,"description":58,"date":418,"categories":58,"tags":348,"draft":206},"\u002Fposts\u002Fmathematical-foundations-of-llms-1","LLM的数学基础（1）：","2026-09-19 22:59:55",{"path":420,"title":421,"description":58,"date":422,"categories":58,"tags":348,"draft":206},"\u002Fposts\u002Ftowards-maximize-serving-capatity","如何在 offline bench 里最大化 LLM serving system 的吞吐","2026-09-19 05:45:05",{"path":424,"title":425,"description":426,"date":427,"categories":382,"tags":428,"draft":311},"\u002Fposts\u002Fllm-memory-compute-tradeoff","LLM 的 Memory-Compute Trade-off：从 MHA 到 MLA、Sparse 与 Linear Attention","ds4.1f 上线了。它的架构 novel 在哪，便宜和快又是从哪省出来的。","2026-09-11 12:00:00",[319,429,430,431,432],"Attention","MLA","DeepSeek","长上下文",{"path":434,"title":435,"description":436,"date":437,"categories":382,"tags":438,"draft":206},"\u002Fposts\u002Fmarkdown-extensions","Markdown 扩展语法速查","本站 Markdown 支持的扩展语法演示与速查：admonition 提示块、GitHub 仓库卡片、spoiler 遮罩、ABC 乐谱、GP 乐谱、音频播放器、浏览器可运行代码块。","2026-09-01 15:30:00",[439],"博客",{"path":441,"title":442,"description":443,"date":444,"categories":341,"tags":445,"draft":206},"\u002Fposts","为什么还需要手写博客？","AI这么强，还要古法写博客吗？","2026-09-01 13:17:53",[439],{"path":447,"title":448,"description":58,"date":449,"categories":58,"tags":348,"draft":206},"\u002Fposts\u002Fllm-serving-1","LLM Serving（一）:以在H200上部署GLM-5.2为例","2026-07-21 06:15:44",{"path":451,"title":452,"description":58,"date":453,"categories":327,"tags":348,"draft":311},"\u002Fposts\u002Finteresting-music-rhythm","有趣的音乐节奏","2026-07-16 12:47:24",{"path":455,"title":456,"description":457,"date":458,"categories":309,"tags":459,"draft":311},"\u002Fposts\u002Ftowards-monosemanticity","LLM 可解释性（一）：SAE与单语义","读 Anthropic《Towards Monosemanticity》。从难以解释的神经元出发, 讲清superposition为什么让单个神经元混成一团, SAE怎么把多语义的激活分解成单语义特征。","2026-07-12 22:00:00",[319,460,461],"可解释性","SAE",{"path":463,"title":464,"description":465,"date":466,"categories":341,"tags":467,"draft":206},"\u002Fposts\u002Fpiketty-ai-china-k-shaped","用皮凯蒂的镜头看 AI 与中国 K 型社会","一次凌晨的对话整理。从 r>g 出发，看 AI 时代的资本分配、中国的世袭化趋势、体制内的二元劳动力市场，以及一个普通 00 后在一线打拼的真实牌面。","2026-06-24 03:00:00",[413,468,357],"经济",{"path":470,"title":471,"description":58,"date":472,"categories":58,"tags":473,"draft":206},"\u002Fposts\u002Fclaude-code-compact-behavior","cluade code compact 行为","2026-06-23 20:17:42",[474,475,476],"Agent","AI Agent","Claude Code",{"path":478,"title":479,"description":58,"date":480,"categories":309,"tags":481,"draft":206},"\u002Fposts\u002Fclaude-code-settings","Claude Code 配置分享","2026-06-23 14:22:12",[319,475,474,476],{"path":483,"title":484,"description":485,"date":486,"categories":382,"tags":487,"draft":311},"\u002Fposts\u002Fmy-thoughts-on-the-ai-industry","简单聊聊对 AI Agent 行业的理解","写这篇文章的初衷，是我发觉逐渐忘了前两个月的 AI\u002FLLMs 是什么样的能力和情况了。意识中似乎某些在突然加速，但理性和直觉告诉我这不对，作为行内人，我知道这是一个渐进的、累积的过程。我想自己回忆一下，Agent的发展。","2026-06-20 13:14:25",[319,475,474],{"path":489,"title":490,"description":58,"date":491,"categories":309,"tags":492,"draft":311},"\u002Fposts\u002Frotational-positional-encoding","RoPE","2026-06-17 18:31:40",[319,318,490],{"path":314,"title":5,"description":58,"date":310,"categories":309,"tags":494,"draft":311},[318,319],{"path":496,"title":497,"description":58,"date":498,"categories":382,"tags":499,"draft":206},"\u002Fposts\u002Fcoding-agent","关于 Coding Agent 使用的遐思","2026-06-03 17:54:20",[474],{"path":501,"title":502,"description":503,"date":504,"categories":309,"tags":505,"draft":206},"\u002Fposts\u002Fabout-ssh-usage","关于SSH的使用——客户端篇","介绍SSH客户端的用法。","2025-06-17 02:44:57",[506],"工具",{"path":508,"title":509,"description":510,"date":511,"categories":382,"tags":512,"draft":311},"\u002Fposts\u002Fopensourcesoftwarestarttutorial","如何开始一个开源项目？","浅谈我对开源项目的认知。","2025-04-17 18:00:00",[513],"开源",{"path":515,"title":516,"description":517,"date":518,"categories":519,"tags":520,"draft":206},"\u002Fposts\u002Fmotivation","渴求动力","有意识的生活并主动地行动会使得生活更美好。","2025-03-05 00:44:57","生活-情感",[413],{"path":522,"title":523,"description":524,"date":525,"categories":519,"tags":526,"draft":206},"\u002Fposts\u002F2024summary","2024年终总结","啊，痛彻的时光；啊，悲怆的人！","2025-01-01 01:00:00",[413],{"path":528,"title":529,"description":530,"date":531,"categories":309,"tags":532,"draft":206},"\u002Fposts\u002Fvps-crash-startup","快速搭建个人VPS","如何快速搭建个人VPS。","2024-12-31 19:00:00",[533],"VPS",{"path":535,"title":536,"description":537,"date":538,"categories":309,"tags":539,"draft":206},"\u002Fposts\u002Fdata-label-software","AI数据（自动）标注软件概述","对市面上目前的AI数据（自动）标注软件进行调研。","2024-12-20 14:00:00",[540,541],"数据标注","深度学习",{"path":543,"title":544,"description":545,"date":546,"categories":519,"tags":547,"draft":311},"\u002Fposts\u002Fself-destruction","生活的崩坏与自我毁灭","是什么在牵引着我？我的行为，我的思想？我的感受？我所处的地方？它们在何处？","2024-12-19 12:00:00",[413],{"path":549,"title":550,"description":551,"date":552,"categories":309,"tags":553,"draft":206},"\u002Fposts\u002Fcs194-196_llmagents","LLM Agents 课程笔记","UCB CS294\u002F194-196 Large Language Model Agents 的课程笔记。","2024-11-30 20:00:00",[474],{"path":555,"title":556,"description":557,"date":558,"categories":519,"tags":559,"draft":206},"\u002Fposts\u002Fwhatsyourworth","价值、位置","你们的生存究竟有何价值？如果毫无价值，你们究竟为何存在？","2024-11-29 02:00:00",[413],{"path":561,"title":562,"description":563,"date":564,"categories":519,"tags":565,"draft":311},"\u002Fposts\u002Fblogrelifever2","重启博客v2","忙里偷闲学习前端，还是决定结合想法手搓一个自己的博客。","2024-11-25 00:00:00",[439],{"path":567,"title":568,"description":569,"date":570,"categories":519,"tags":348,"draft":206},"\u002Fposts\u002Fblogrelife","重启博客","在我的表达欲与现实所需(问题解决记录)驱动下，我决定重写博客。","2024-10-21 20:44:57",{"path":572,"title":573,"description":574,"date":575,"categories":341,"tags":576,"draft":311},"\u002Fposts\u002Fsurvival-welcome-to-university","生存手册——欢迎来到大学","大学同时生产三样东西：可验证技能、认证信号和驯化资本。你付了四年的钱，真正拿到手的是哪一样，取决于你毕业时要进的那个市场在为什么付费。","2024-10-12 20:00:00",[577,578,579],"大学","求职","方法论",{"path":581,"title":582,"description":583,"date":584,"categories":519,"tags":585,"draft":311},"\u002Fposts\u002Frevival","240205 复活","好吧，断更了许久，我又终于回来了，我又终于回来了。","2024-02-05 20:00:00",[413,586,587],"生活","返乡",{"path":589,"title":590,"description":591,"date":592,"categories":519,"tags":593,"draft":206},"\u002Fposts\u002Fclouds","「230717」云","我已经很久忘记了不我已经很少把自己的想法在lofter上发布的。","2023-07-17 20:00:00",[413,594],"散文",1795115926099]