搜索与文字目录 / Search

让输入直接绕过一层与输出相加，是深层网络能训起来的结构性前提。

梯度在层层回传中不断相乘，指数级衰减到接近零，导致浅层学不动。

模型规模、数据量、算力同步放大时，损失会以可预测的幂律下降。

在海量文本上训练出的超大 Transformer，本质是在预测下一个 token。

模型看不见字符，只看见 token —— 文本被切成的一个个小片段。

把文本变成一串数字（向量），让语义相近的内容在向量空间里也靠得近。

模型单次能「看见」的 token 总量上限——它的工作台有多大。

放在上下文中段的信息，模型的利用率明显低于放在开头和结尾。

控制模型从「下一个词的概率分布」里怎么挑的一组旋钮。

在预训练模型的基础上，用特定数据继续训练，改变模型的权重。

沿着「误差下降最快的方向」一小步一小步调整参数，是模型训练的基本方法。

高效计算「每个参数该往哪调」的算法，让深层网络的训练成为可能。

模型把训练数据背下来了，包括其中的噪声，导致遇到新数据就失灵。

用滑动的小窗口在图像上提取局部特征的网络，长期是计算机视觉的主力。

按顺序逐个处理序列、并把「记忆」传给下一步的网络，Transformer 之前的序列主力。

用带标准答案的数据训练模型，让它学会从输入预测输出。

从没有标注的数据中自己找出结构与规律。

让智能体在环境中试错，用奖励信号学出一套行为策略。

从数据本身构造训练目标，不需要人工标注——大模型时代的直接前提。

一整套限制模型「死记硬背」能力的手段，用来对付过拟合。

用一连串「是/否」判断做预测的模型，以及把多棵树组合起来的增强方法。

在没有标签的情况下，把相似的样本自动分成若干组。

把高维数据压缩到低维，同时尽量保留原有的结构信息。

维度越高，空间越空旷，「距离」越失去区分度——高维空间不符合直觉。

在海量无标注文本上做自监督训练，是大模型能力的主要来源。

回答前先去知识库检索相关资料，连同问题一起交给模型，让它基于材料作答。

专门存储向量并支持「找出最相似的 K 个」这种查询的数据库。

通过设计输入的措辞与结构，让模型稳定产出你想要的结果。

把长文档切成小块，让检索能定位到片段而不是整本书。

根据问题从知识库里找出最相关的内容——RAG 效果的真正瓶颈。

对召回的候选逐条精算相关性并重新排序，是「先粗筛后精排」的第二步。

在生成的每一步屏蔽掉会破坏格式的 token，从机制上保证输出合法。

把重复出现的提示前缀的中间计算结果缓存下来，省掉重复计算。

让模型稳定吐出程序能直接解析的格式，而不是一段自然语言。

让模型把推理过程写出来再给答案，复杂任务上准确率明显提升。

能自主循环「思考→调用工具→看结果→再思考」直到完成任务的 LLM 系统。

让模型输出结构化的「调用意图」，由外部程序实际执行并把结果回传。

把「模型如何连接外部工具与数据」标准化的开放协议。

MCP 的三个角色（Host / Client / Server）与两种传输方式。

Agent 反复执行的四步：接收输入 → 推理规划 → 调用工具 → 观察结果。

让 Agent 跨轮次、跨会话保留信息的机制——上下文窗口装不下的部分。

通过学习「从噪声一步步去噪」来生成内容的模型，当今图像生成的主流。

根据文本描述生成图像，当前主要由扩散模型驱动。

让一个模型同时处理文本、图像、音频等不同形式的信息。

让生成器和判别器对抗训练来生成内容，扩散模型之前的图像生成主力。

语音转文字（识别）与文字转语音（合成），两个反向的任务。

根据文本或图像生成连续视频，难点在于时间一致性。

在纯文本提示之外，用额外条件精确控制生成结果。

系统地衡量一个 LLM 应用好不好——没有它，一切优化都是盲调。

在基础 RAG 之上加一层处理，解决「检索不到」和「检索不够」的问题。

让模型跑得更快更省的一组技术，大多围绕 KV 缓存和注意力。

用更低精度存储模型权重，大幅减小体积和显存，换取轻微的质量损失。

把模型跑起来对外服务的几种方式，各有隐私、成本、可控性的取舍。

用人类偏好训练模型倾向于「人更想要的」回答——把基座模型调教成助手的关键一步。

主动扮演攻击者去攻破自己的模型，在上线前找出漏洞。

试图搞清楚模型内部到底在做什么——目前仍是开放的研究前沿。

把大模型拆成很多「专家」，每个 token 只激活一小部分，参数量大但计算省。

用大模型当老师训练小模型，让小模型学到接近大模型的能力。

用模型生成训练数据，应对真实高质量数据见底的问题。

当前主要的大模型家族及其定位——这是变化最快的一层。

模型生成了流畅、自信，但事实上错误或纯属虚构的内容。

把恶意指令藏在模型会读到的内容里，劫持它的行为。

用特定措辞诱导模型绕过自身的安全限制，说出本不该说的内容。

系统找到了让奖励变高、却违背设计者本意的捷径——它没做错，是目标没写对。

让模型的行为符合人类意图与价值观的一整套技术与目标。

在回答前先生成一段长思考过程的模型，用推理时的算力换准确率。

帮你搭 Agent 循环、状态管理和多 Agent 编排的工具库。

把工作拆给多个 Agent 协作完成，两种主流模式：中心调度与对等交接。

在模型前后加多道独立检查，用分层防御把风险挡在外面。

不是所有任务都需要最强的模型——先用最强的建立基线，再往下换。

模型只靠提示里的几个例子就学会新任务，不改一个权重。

决定每次调用时上下文窗口里到底放什么——比雕琢单条提示更上层的问题。

让模型交替进行「推理」和「调用工具」，是 Agent 循环的经典范式。

在训练数据里掺入恶意样本，让模型学到攻击者想要的行为。

对输入做人眼难察的微小扰动，就能让模型给出完全错误的结果。

模型会继承并放大训练数据里的社会偏见，在关键决策场景造成不公。

模型可能记住并泄露训练数据，加上数据出境等合规约束，是落地的硬门槛。

管理 AI 风险的制度、标准与法律——技术之外，决定 AI 怎么被允许使用。

模型一边生成一边把 token 逐个吐出来，而不是等全部生成完再返回。

模型每步输出的 token 概率，可用来估计它对自己回答有多确定。

在对话最前面设定模型的身份、规则和边界，贯穿整场对话。

同一问题让模型独立解多次，取多数答案，比单次思维链更可靠。

让模型像搜索一样展开多条推理分支、评估、回溯，而不是一条道走到黑。

记录 LLM 应用每次调用的全过程，让你能看清、能调试、能算账。

让模型写代码并真正跑起来，用执行结果解决它自己算不准的问题。

让 Agent 像人一样看屏幕、点鼠标、敲键盘，操作没有 API 的软件。

让模型检查并批判自己的输出，再据此改进——用一轮自我审视换质量。

把知识组织成实体和关系的网络，让检索能沿关系走、能做全局归纳。

用海量图文配对训练，让图片和描述它的文字在向量空间里对齐。

把数据压进一个连续的潜在空间再重建，生成模型三巨头之一。

生成音乐、音效、环境声——和语音合成是不同的任务。

把低清图放大、把破损图修好——用生成能力补出缺失的细节。

生成的不只是一张图或一段视频，而是可交互、有物理规律的三维世界。

判断一张图/一段文/一段音是不是 AI 生成的——生成能力越强，越难也越必要。

让模型写代码——从补全一行到实现整个功能，AI 最成功的应用之一。

把一个大目标拆成可执行的小步骤——Agent 能不能办成复杂任务的关键。

在关键或不可逆的步骤上让人介入确认，是 Agent 安全落地的现实底线。

把代码生成、执行、仓库理解整合进开发流程的工具，这一层变化极快。

用预先定义好的固定流程串起多次 LLM 调用，和自主 Agent 是两条路。

让模型对照一套写下来的原则自我批判、自我修正，用 AI 反馈替代大量人工标注。

把助手回复的开头先替它写好，强力约束输出的格式和走向。

让模型的回答带出处、可核实——把「信不信它」变成「查一查」。

长任务里自动把旧上下文摘要、裁剪，腾出窗口又不丢关键信息。

把一套指令、工具、知识打包成可复用、可按需加载的「技能」，供 Agent 调用。

训练时按小批量的统计量把每层输入归一化到稳定分布，让深层网络训得更快、更稳。

把「预测离目标有多远」压成一个数，训练就是不断调参数把这个数变小。

用「熵」量化不确定性与信息量，交叉熵、KL 散度是训练与评估语言模型的基本度量。

用「核技巧」在高维空间找最大间隔的分界面，深度学习兴起前最强的分类器之一。

在预训练之后，用指令、偏好与可验证反馈把基座模型塑造成可用助手。

把顺序和相对距离注入注意力，否则 Transformer 无法区分 token 的先后。

按单个样本的特征维稳定激活尺度，是深层 Transformer 能稳定训练的关键组件。

决定拿到梯度后怎样更新参数、每一步走多远，直接影响训练速度与稳定性。

冻结大部分基础权重，只训练少量适配参数，以更低显存和存储成本定制模型。

把模型、数据和计算拆到多张加速卡上，突破单卡显存与训练时间限制。

把相关样本表示拉近、不相关样本推远，学出可检索、可迁移的表示空间。

在不重新完整训练的情况下组合多个模型或适配器的能力，但容易产生参数干扰。

用可复现任务衡量模型能力、鲁棒性与安全性，并识别污染和指标错配。

按请求难度、成本与风险选择不同模型，让系统不必所有任务都调用最贵模型。

上线后的输入、用户与目标会变化，持续监控负责发现模型质量随环境悄悄退化。

明确 Agent 代表谁、能访问什么、凭什么执行，并隔离密钥与高风险权限。

直接学习把噪声连续运输到数据的速度场，是与扩散紧密相关的现代生成路径。

管理训练数据的来源、许可、质量、去重、配比与可追溯性，决定能力也决定风险。

让置信度与真实正确率匹配，并在风险过高时拒答、检索或转人工。

在回答阶段投入更多候选、搜索、验证和工具计算，以成本与延迟换取成功率。

用压缩状态线性处理长序列，并通过选择机制决定哪些信息保留或遗忘。

从目标说话者的参考语音提取身份表示或声学提示，生成该说话者从未说过的新语音；可在推理时零样本条件化，也可用较长语音微调模型提高一致性。

从查询改写、混合召回、重排、多跳、纠错检索和答案引用理解何时需要升级基础 RAG。

从决策边界、FGSM/PGD、鲁棒优化和认证半径，到自适应攻击、分布外变换与系统级损失控制，理解保证的适用域。

从「一问一答」到「给个目标、自己一路做完」

从模型适配、工具注册、状态图、持久化和追踪，到抽象泄漏、版本迁移与逃生口，判断何时框架值得引入。

模型提出意图，受控执行层证明谁能对什么做哪件事

把一次回答变成可观察、可校验、能停下来的状态机

把会话上下文、工作状态、情景记录和长期事实分开，理解写入、整合、检索、遗忘、冲突与隐私治理。

从触发描述、渐进披露、脚本与资产，到选择混淆、权限、版本和回归，理解技能与提示、工具和工作流的边界。

让模型的行为，符合人类的意图与价值观

处理每个词时，让模型自己决定该「看」句子里的哪些词

从采样率、神经编解码器和多码本token，到自回归/扩散生成、文本与旋律条件、长程音乐结构及主观评测。

从计算图、局部导数和上游梯度开始，完整手算一次反向遍历，并理解分叉累加、向量—雅可比积、激活缓存和梯度检查。

从卷积张量的统计轴、训练/推理双路径、运行均值到同步 BatchNorm，理解它为何加速许多视觉网络，又为何会在小批量、域迁移和部署融合时失效。

用同一招聘案例理解代表性偏差、标签偏差、代理变量、人口统计平等、均等机会、校准及其不可兼得边界。

理解固定长度、结构、语义与父子切分，处理重叠、上下文丢失、表格代码和嵌入预算。

从原子主张、稳定锚点和蕴含判断，到引用正确性、完整性与来源质量，建立真正可核验的回答。

从双编码器、归一化相似度矩阵和对称交叉熵，到零样本分类、检索、组合关系与偏差边界。

从 K-means 的分配—更新手算，到层次、DBSCAN、混合模型、选 k、稳定性和业务解释。

From manual K-means assignment–update computation to hierarchical clustering, DBSCAN, mixture models, choosing k, stability, and business interpretation.

从离散卷积、权重共享和输出形状，到感受野、采样混叠、平移等变与现代视觉骨干。

理解生成—执行—观察闭环的能力来源，以及进程、文件、网络、资源、凭据与人工授权组成的隔离边界。

让模型写代码，从补全一行到实现整个功能

从搜索、调用链定位和最小编辑，到测试、差异审查、工作树保护与安全回退，理解仓库级智能体为何不只是代码生成。

从截图、DOM与可访问性树的动作接地，到焦点、幂等、权限和恢复，理解为什么“会点按钮”远不等于可靠完成任务。

从原则文本、情境解释、冲突优先级到 AI 反馈训练，理解怎样减少逐例人工标签，以及为什么原则选择仍然是治理问题。

From principle text, contextual interpretation, and conflict priority to AI feedback training, understand how to reduce per-example human labels and why principle selection remains a governance issue.

从 JSON Schema、文法状态和 tokenizer 边界，到死路、复杂度、流式输出与语义校验。

从基率、ROC和校准，到模型指纹的分布漂移、鲁棒水印、C2PA式签名清单、链式编辑与“无信号不等于真人创作”。

理解截断、摘要、抽取、检索与提示压缩的有损本质，并设计不可丢约束、来源追踪和压缩回归。

把提示词、检索证据、工具结果、记忆与历史消息当成有限预算下的信息系统，而不是把所有文本一股脑塞进窗口。

模型一次能「看见」的 token 总量上限

从 InfoNCE、温度与梯度方向，到增强不变性、大批量、假负样本、表示坍缩和迁移验收。

从条件概率和classifier-free guidance，到边缘、深度、姿态、分割、参考图、多条件冲突、控制强度与可验证控制。

让模型先写出一步步推理，再给答案——复杂题上更准

从体积指数增长、距离集中和样本复杂度理解高维检索、密度估计与降维的必要性。

Understanding the necessity of high-dimensional retrieval, density estimation, and dimensionality reduction through exponential volume growth, distance concentration, and sample complexity.

从 P(x)、P(y) 与 P(y|x)，到 PSI、延迟标签、代理指标、切片告警和安全响应。

区分可用性破坏、定向错分、干净标签攻击和后门，从污染率与梯度影响一路追到数据谱系、训练检测和触发器评测。

从不纯度、信息增益、剪枝到随机森林和梯度提升，理解树模型对表格数据的优势与泄漏风险。

From impurity, information gain, and pruning to random forests and gradient boosting, understand the advantages of tree models for tabular data and the risk of leakage.

从运行位置、排队与批处理，到版本发布、幂等副作用、降级和事故响应。

学习「从一团噪声一步步去噪」，把图像「显影」出来

从 PCA、随机投影、t-SNE、UMAP 与自编码器理解压缩目标、可视化失真、数据泄漏和下游评测。

Understand compression objectives, visualization distortion, data leakage, and downstream evaluation through PCA, random projection, t-SNE, UMAP, and autoencoders.

从软标签、温度与 KL，到 token/序列/特征蒸馏、覆盖缺口、容量上限和偏差继承。

从全局批量与 all-reduce，到张量并行、pipeline bubble、ZeRO/FSDP、3D 并行和故障检查点。

把文字变成向量，让「意思相近」变成「距离相近」

把业务目标拆成可观察属性，定位检索、生成、工具与流程各自造成的失败。

在预训练的「通才」上继续训练，调成某个任务的「专才」

从概率路径、条件速度和边缘速度，到直线路径、ODE采样、数值误差、配对设计及与扩散/连续流的关系。

从极小极大博弈、最优判别器和JS散度，到非饱和损失、模式坍塌、训练振荡、Wasserstein距离与评测边界。

从系统清单、风险分级和责任矩阵，到上线门、变更审批、事件响应与退役，理解治理如何把原则落到生命周期证据。

从一维斜率和方向导数出发，手算一次更新，再看学习率、曲率、mini-batch、动量与 Adam 为什么改变训练轨迹。

从输入、上下文、输出、工具和运行时五层控制，到阈值成本、策略组合和故障注入，理解护栏能保证什么、又不能保证什么。

流畅、自信、却是编的——为什么这是结构性的，而不是「再大一点就好」

从风险分层、升级阈值、审核界面和队列容量，到自动化偏见与反馈采样，设计真正降低伤害的人机闭环。

把一句话变成一张没人拍过的图

不改一个参数，只靠提示里的几个例子就学会新任务

从 TTFT/TPOT、算术强度和连续批处理，到 PagedAttention、FlashAttention、量化、推测解码与真实负载验收。

先分清真实分布 P 与模型分布 Q，再从自信息、熵逐步推出交叉熵、KL 散度和困惑度。

比较特征归因、探针、反事实、激活替换、稀疏特征和电路分析，理解忠实度、稳定性、完备性与“可读不等于真实”。

从直接/间接、多轮、编码、后缀优化与多模态越狱，到威胁模型、基率、纵深防御、自适应红队和正常请求误拒。

从最大间隔、对偶问题与核技巧理解线性、RBF、多项式核的能力、超参数和规模限制。

从本体、实体解析和时态边，到路径检索与 GraphRAG，理解图何时优于相似度、又怎样传播错误。

从「预测下一个词」到「像助手一样对话」

从 token 对数概率、序列似然、校准和选择性预测理解何时 logprob 有用，以及分词与语义等价如何误导阈值。

从业务代价到可微代理，从单样本误差到经验风险；用同一个数值例子看见损失、梯度和参数更新怎样接成闭环。

理解长上下文中的位置效应，学会用受控实验测量它，并用检索、重排和结构设计降低关键证据被忽略的概率。

让任何工具和数据，都能用同一个标准接口插进 AI 应用

从 Host–Client–Server、JSON-RPC、能力协商与三类原语，理解 MCP 标准化了什么，以及授权与执行安全仍由谁负责。

从估计对象、采样预算和 pass@k，到污染、裁判偏差、置信区间与发布门槛。

不背品牌榜单，而是用架构、目标、模态与产品层四个轴，判断一个模型为什么擅长某类任务。

从检查点平均、模型汤和任务向量，到置换对齐、符号冲突、TIES/DARE、系数搜索与安全回归。

从直接路由、先小后大和质量差预测，到升级阈值、级联延迟、公平风险与在线再校准。

从任务分布、强模型上界和多维门禁，到每成功任务成本、帕累托前沿与退出策略。

从 top-k 路由与加权输出，到负载均衡、容量溢出、all-to-all、专家专化和推理批处理。

从任务依赖图、通信拓扑和共享状态，到关键路径、移交损失与同源错误，判断何时多智能体真的优于单智能体。

让一个模型同时看懂图、听懂声、读懂字

从 LayerNorm 的中心化、RMSNorm 的均方根缩放，到 Pre-Norm/Post-Norm 与实现验收，理解归一化究竟沿哪条轴统计、保留什么、又丢掉什么。

用 trace、span、版本、质量信号和最小化数据，把“偶发答错”变成可定位、可回放、可回归的问题。

从 SGD、Momentum、Adam 与 AdamW 的状态更新，到预热、平台、衰减、批量规模和混合精度；理解一套训练配方为什么必须整体调试。

从经验风险与未知风险的分离开始，读懂训练/验证曲线，识别泄漏、重复与调参污染，并理解容量、双下降和大模型记忆的边界。

从 ΔW=BA、rank 与缩放，到可训练参数账、QLoRA、目标模块、合并版本和多适配器干扰。

从目标、前置条件、产物和完成谓词理解计划—执行分离、滚动规划、检查点与局部重规划。

从排列等变、正弦绝对位置、相对偏置到旋转位置编码，理解顺序怎样进入注意力，以及长上下文扩展为何不是改一个窗口数字。

预训练给能力，后训练决定这些能力怎样被调用和呈现

理解 assistant prefix 如何改变条件分布，并与输入 prefill、提示缓存和约束解码严格区分。

在海量无标注文本上自监督学习，一次性把「通用能力」灌进模型

从采集、提示、检索、日志、供应商到模型权重，逐站理解个人数据怎样扩散，以及怎样用最小化、隔离和可验证删除缩小风险。

从最长共同 token 前缀、KV 占用与命中价值，到提示排序、路由亲和、失效和租户隔离。

设计输入的措辞与结构，让模型稳定产出你想要的结果

把恶意指令藏在模型会读到的内容里，劫持它的行为

从 scale、zero-point 和舍入误差，到粒度、异常值、PTQ/QAT、权重/激活/KV 与硬件内核。

回答前先检索相关资料，让模型「开卷作答」而不是凭记忆瞎编

用外部证据修正假设，而不是让一条推理链凭空走到底

把更多训练与推理时算力用于搜索、验证和修正，换取难题准确率

从资产、攻击者和攻击面，到测试用例、证据、修复回归与剩余风险，理解红队为何不是收集几条越狱提示。

区分空泛自评、外部反馈和可验证修正，用最小改动、停止条件与版本化经验避免“再想一遍”制造更自信的错误。

从 L2/L1、AdamW、Dropout、数据增强与早停出发，理解参数、表示、数据和训练路径四类约束怎样改变泛化，而不是把正则化当成一个万能旋钮。

用 MDP、回报、价值函数、Bellman 方程、探索与策略梯度理解 Agent 如何在交互中优化长期结果。

理解双塔召回、交叉编码器、late interaction、LLM 重排和位置偏差，并设计 Recall@k 到端到端答案的评测。

从 y=x+F(x) 的前向增量与 I+JF 的反向雅可比开始，理解深度退化、投影捷径、分支尺度、零初始化以及 Transformer 的 Pre-Norm/Post-Norm。

从知识库里找出最相关的少量内容，喂给模型作答

从目标与代理指标的缝隙出发，识别规格博弈、评审器过拟合和反馈篡改，并用独立评测与分层防线降低风险。

从示范数据、偏好比较、奖励模型到策略优化，理解 RLHF 在优化什么，以及它为什么不能等同于“让模型绝对安全”。

从共享递推、时间展开与 BPTT，到梯度连乘、LSTM 门控、教师强制和流式部署。

用一组可手算的候选分布串起 temperature、top-k、top-p、重复惩罚、停止条件与随机种子，理解每个旋钮改变哪一步、不能保证什么。

从 log-log 直线、边际收益递减到计算最优分配，再到数据质量、能力阈值与全生命周期成本。

理解采样多样性、答案聚合、相关错误、成本与置信估计，并区分多数一致和事实正确。

统一理解自回归、遮盖重建、对比学习与自蒸馏，并追踪预文本目标怎样迁移、走捷径或错位。

把语音识别、语音合成和声音克隆放到同一管线，理解声学特征、对齐、声码器、流式延迟、WER及身份授权。

从连续状态方程、离散化与卷积等价，到选择性扫描、线性复杂度、稳定性和随机访问边界。

从 token、UTF-8 字节块与 SSE 事件，到 TTFT、背压、结构缓冲、安全审查和最终提交。

从最小 schema、约束生成和分层校验，到版本迁移、有限修复与安全执行。

从模糊—降采样—噪声模型、像素损失与感知损失，到GAN/扩散先验、盲超分、幻觉边界和面向任务评测。

从「输入 + 目标」样本中学习，并在未见数据上做出可靠预测

从教师示范、程序真值、自训练和增强，到验证器选择偏差、覆盖配比、反馈回路与模型坍缩。

从消息序列、指令优先级、冲突解析和上下文拼装，到注入、泄露、版本化与契约测试，理解系统提示能控制什么。

不改权重，靠采样、搜索、工具与检查把额外计算换成更可靠的答案

从 Unicode、预切分、BPE/Unigram、字节回退到特殊 token，理解同一句文本为何会变成不同长度、不同成本和不同能力边界。

给只会说话的模型，接上一双能查、能算、能做事的「手」

从来源与授权、谱系、过滤去重和配比，到污染、删除请求、训练快照与事故响应。

理解状态、候选生成、价值评估、BFS/DFS、剪枝和搜索预算，并区分探索空间扩大与答案正确。

从置信分数、可靠性图和 ECE，到选择性预测、拒答阈值与分布漂移。

用聚类、降维、密度估计和生成建模理解“没有标准答案”的目标设计、非唯一性、伪结构与验证。

从编码分布、重参数化和ELBO，到KL—重建权衡、后验坍塌、潜空间插值和生成质量边界。

从标量连乘到矩阵雅可比，从 sigmoid 饱和到 RNN 时间展开；用数值例子理解初始化、门控、残差、归一化和梯度裁剪分别改变哪段路径。

理解向量、距离度量、HNSW/IVF/PQ、过滤、更新与一致性，并把索引 Recall、延迟和端到端 RAG 分开。

从视频潜变量、时空注意力和扩散，到图生视频、级联超分、身份一致、物理失败、长视频规划与时序评测。

把“说什么”和“像谁说”拆成两种条件，再用可懂度、自然度、身份相似度与授权共同验收

从 DAG、状态机和持久执行，到幂等、重试、补偿、版本迁移与人工任务，建立可恢复、可审计的 AI 流程。

从状态表示、转移与奖励模型，到想象滚动、模型预测控制、不确定性和模型利用，理解“在脑中试走”怎样帮助又怎样误导智能体。

From shared recurrence, time unrolling, and BPTT, to gradient multiplication, LSTM gating, teacher forcing, and streaming deployment.

From discrete convolution, weight sharing, and output shapes, to receptive fields, sampling aliasing, translation equivariance, and modern visual backbones.

Turn text into vectors, making “similar meaning” become “close in distance”

When processing each word, let the model decide for itself which words in the sentence to 'look at'.

Starting from computational graphs, local derivatives, and upstream gradients, work through one complete backward traversal by hand, and understand branch accumulation, vector–Jacobian products, activation caching, and gradient checking.

From Unicode, pre-tokenization, BPE/Unigram, byte fallback, and special tokens, understand why the same text yields different lengths, different costs, and different capability boundaries.

From the statistical axes of convolutional tensors, training/inference dual paths, and running means to synchronized BatchNorm, understand why it accelerates many visual networks and why it fails with small batches, domain shift, and deployment fusion.

Continue training on a pre-trained “generalist” to turn it into a task-specific “specialist”

From soft labels, temperature, and KL, to token/sequence/feature distillation, coverage gaps, capacity limits, and bias inheritance.

The maximum total number of tokens a model can “see” at once

From ΔW=BA, rank, and scaling, to trainable parameter accounting, QLoRA, target modules, merged versions, and multi-adapter interference.

Self-supervised learning on massive unlabeled text, instilling “general capabilities” into the model in one go

From scale, zero-point, and rounding error, to granularity, outliers, PTQ/QAT, weights/activations/KV, and hardware kernels.

Pre-training provides capabilities; post-training determines how those capabilities are invoked and presented.

From LayerNorm's centering and RMSNorm's root-mean-square scaling, to Pre-Norm/Post-Norm and implementation acceptance, understand exactly which axis normalization computes statistics along, what it retains, and what it discards.

From the state updates of SGD, Momentum, Adam, and AdamW, to warm-up, plateau, decay, batch size, and mixed precision; understand why a training recipe must be tuned as a whole.

From business cost to a differentiable proxy, from single-sample error to empirical risk; using the same numerical example, see how the loss, gradient, and parameter update connect into a closed loop.

A unified understanding of autoregression, masked reconstruction, contrastive learning, and self-distillation, and tracing how pretext objectives transfer, take shortcuts, or become misaligned.

From permutation equivariance, sinusoidal absolute positions, and relative bias to rotary position embeddings, understand how order enters attention and why long-context extension is not simply changing a window number.

From scalar products to matrix Jacobians, from sigmoid saturation to RNN unrolling over time; use numerical examples to understand how initialization, gating, residuals, normalization, and gradient clipping each change which segment of the path.

From InfoNCE, temperature, and gradient direction, to augmentation invariance, large batches, false negative samples, representation collapse, and transfer validation.

From dual encoders, a normalized similarity matrix, and symmetric cross-entropy, to zero-shot classification, retrieval, compositional relations, and bias boundaries.

From message sequences, instruction priority, conflict resolution, and context assembly to injection, leakage, versioning, and contract testing—understand what system prompts can control.

Treat prompts, retrieved evidence, tool results, memory, and historical messages as an information system under a limited budget, rather than cramming all the text into the window.

Understand how the assistant prefix changes the conditional distribution, and strictly distinguish it from input prefill, prompt caching, and constrained decoding.

From minimal schema, constraint generation, and layered validation to version migration, limited repair, and secure execution.

From task distribution, strong-model upper bound, and multidimensional gates, to cost per successful task, Pareto frontier, and exit strategy.

Surface the most relevant snippets from a knowledge base and feed them to the model for answering.

Use traces, spans, versions, quality signals, and minimized data to turn “occasional wrong answers” into locatable, replayable, regression-testable problems.

From P(x), P(y), and P(y|x) to PSI, delayed labels, proxy metrics, slice alerts, and safe response.

Fluent, confident, but fabricated — why this is structural, not 'just make it bigger'.

From collection, prompts, retrieval, logs, vendors, to model weights, understand station by station how personal data spreads, and how to reduce risk through minimization, isolation, and verifiable deletion.

Distinguish vague self-assessments, external feedback, and verifiable corrections; use minimal changes, stopping conditions, and versioned experience to avoid “thinking it over again” producing more confident errors.

Understand sampling diversity, answer aggregation, correlated errors, cost, and confidence estimation; and distinguish majority agreement from factual correctness.

Understand planning–execution separation, rolling planning, checkpoints, and local replanning through goals, preconditions, artifacts, and completion predicates.

From model adaptation, tool registration, state graphs, persistence, and tracing, to abstraction leaks, version migration, and escape hatches, determine when a framework is worth adopting.

Separate conversational context, working state, episodic records, and long-term facts; understand writing, integration, retrieval, forgetting, conflict, and privacy governance.

Let models write code, from completing a line to implementing an entire feature.

From task dependency graphs, communication topologies, and shared state to critical paths, handoff losses, and common-source errors, determine when multi-agent truly outperforms single-agent.

From risk stratification, escalation thresholds, review interfaces, and queue capacity to automation bias and feedback sampling, design a human-machine loop that truly reduces harm.

From DAGs, state machines, and durable execution, to idempotency, retries, compensation, version migration, and human tasks, build recoverable, auditable AI processes.

From probability paths, conditional and marginal velocities, to straight-line paths, ODE sampling, numerical error, coupling design, and the relationship with diffusion/continuous flows.

From the encoding distribution, reparameterization, and the ELBO, to the KL–reconstruction trade-off, posterior collapse, latent-space interpolation, and limits of generation quality.

From minimax games, optimal discriminators, and JS divergence, to non-saturating loss, mode collapse, training oscillation, Wasserstein distance, and evaluation boundaries.

From base rates, ROC, and calibration, to distribution drift in model fingerprints, robust watermarking, C2PA-style signature manifests, chained edits, and “no signal does not equal human creation.”

From assets, attackers, and attack surfaces, to test cases, evidence, remediation regression, and residual risk, understand why red teaming is not about collecting a few jailbreak prompts.

Distinguish availability disruption, targeted misclassification, clean-label attacks, and backdoors, and trace from poisoning rates and gradient influence to data lineage, training-time detection, and trigger evaluation.

From decision boundaries, FGSM/PGD, robust optimization and certified radius, to adaptive attacks, out-of-distribution transformations and system-level loss control, understand the applicable domain of guarantees.

Making model behavior align with human intentions and values

Use more training and inference compute for search, verification, and correction in exchange for accuracy on hard problems.

Understand two-tower retrieval, cross-encoders, late interaction, Large Language Model (LLM) reranking, and position bias, and design evaluation from Recall@k to end-to-end answers.

From query rewriting, hybrid retrieval, reranking, multi-hop, corrective retrieval, and answer citation, understand when to upgrade basic RAG.

From conditional probability and classifier-free guidance, to edges, depth, pose, segmentation, reference images, multi-condition conflicts, control strength, and verifiable control.

From sampling rate, neural codecs, and multi-codebook tokens to autoregressive/diffusion generation, text and melody conditioning, long-range musical structure, and subjective evaluation.

Starting from the gap between goals and proxy metrics, identify specification gaming, evaluator overfitting, and feedback tampering, and reduce risk through independent evaluation and layered defenses.

Understand where the generate–execute–observe loop’s capabilities come from, and the isolation boundary made up of processes, files, network, resources, credentials, and human authorization.

From action grounding in screenshots, the DOM, and the accessibility tree, to focus, idempotency, permissions, and recovery, understand why 'knowing how to click buttons' is far from reliably completing tasks.

From JSON Schema, grammar states, and tokenizer boundaries to dead ends, complexity, streaming output, and semantic validation.

From tokens, UTF-8 byte chunks, and SSE events, to TTFT, backpressure, structural buffering, safety review, and final commit.

From runtime location, queuing and batching, to version release, idempotent side effects, degradation, and incident response.

From global batch and all-reduce, to tensor parallelism, pipeline bubble, ZeRO/FSDP, 3D parallelism, and fault checkpointing.

From state representations, transition and reward models, to imagined rollouts, model predictive control, uncertainty, and model exploitation, understand how “trying things out in the mind” helps and misleads agents.

Understand position effects in long contexts, learn to measure them with controlled experiments, and use retrieval, reordering, and structural design to reduce the probability that key evidence is ignored.

Instead of memorizing brand rankings, use four axes—architecture, objectives, modality, and product layer—to judge why a model excels at a particular type of task.

From checkpoint averaging, model soups, and task vectors, to permutation alignment, sign conflict, TIES/DARE, coefficient search, and safety regression.

From top-k routing and weighted outputs, to load balancing, capacity overflow, all-to-all, expert specialization, and inference batching.

Use a set of hand-computable candidate distributions to tie together temperature, top-k, top-p, repetition penalty, stopping conditions, and random seed, and understand which step each knob changes and what it cannot guarantee.

Understand states, candidate generation, value evaluation, BFS/DFS, pruning, and search budget, and distinguish between expanding the search space and getting the answer right.

Using clustering, dimensionality reduction, density estimation, and generative modeling to understand the objective design, non-uniqueness, Pseudo-structure, and validation when there is “no standard answer”.

The model states intent; the controlled execution layer proves who can perform which action on what.

From search, call-chain localization, and minimal edits to testing, diff review, working tree protection, and safe rollback, understand why repository-level agents are more than code generation.

Use external evidence to revise assumptions instead of letting a reasoning chain run its course unsupported

From MCP Host–MCP Client–MCP Server, JSON-RPC, capability negotiation, and three types of primitives, understand what MCP standardizes, and who remains responsible for authorization and execution security.

From trigger descriptions, progressive disclosure, scripts and assets, to selection confusion, permissions, versions, and regressions, understand the boundaries between skills, prompts, tools, and workflows.

From “one question, one answer” to “give it a goal and it gets the whole thing done on its own”

From “predicting the next word” to “conversing like an assistant”

Use the same hiring example to understand representational bias, label bias, proxy variables, demographic parity, equal opportunity, calibration, and the boundaries where they cannot all be satisfied simultaneously.

From demonstration data, preference comparisons, reward models, to policy optimization, understand what RLHF optimizes and why it cannot be equated with “making the model absolutely safe.”

Understand the lossy nature of truncation, summarization, extraction, retrieval, and prompt compression, and design must-not-drop constraints, source tracking, and compaction regression.

Have the model write out step-by-step reasoning before giving an answer—more accurate on complex problems

Starting from the separation of empirical risk and unknown risk, learn to read training/validation curves, recognize leakage, repetition, and tuning contamination, and understand the boundaries of capacity, double descent, and large model memorization.

Use MDPs, returns, value functions, Bellman equations, exploration, and policy gradients to understand how an RL agent optimizes long-term outcomes through interaction.

Starting from one-dimensional slopes and directional derivatives, work through one update by hand, then see why learning rate, curvature, mini-batch, momentum, and Adam change the training trajectory.

One model that understands images, sounds, and text all at once.

Starting from L2/L1, AdamW, Dropout, data augmentation, and early stopping, understand how four types of constraints—on parameters, representations, data, and training paths—change generalization, instead of treating regularization as a one-size-fits-all knob.

Split "what is said" and "who it sounds like" into two conditions, then accept using intelligibility, naturalness, speaker similarity, and authorization together

Understand when logprob is useful from token log probabilities, sequence likelihood, calibration, and selective prediction, and how tokenization and semantic equivalence mislead thresholds.

Design the wording and structure of inputs so that the model consistently produces the results you want.

From estimation target, sampling budget, and pass@k, to contamination, judge bias, confidence intervals, and release thresholds.

From longest common token prefix, KV footprint, and hit value, to prompt ordering, routing affinity, invalidation, and tenant isolation.

From blur–downsampling–noise models, pixel loss and perceptual loss, to Generative Adversarial Network (GAN)/diffusion priors, blind super-resolution, hallucination boundaries, and task-oriented evaluation.

Put speech recognition, speech synthesis, and voice cloning into the same pipeline, understanding acoustic features, alignment, vocoders, streaming latency, WER, and identity authorization.

Give a model that can only talk a pair of “hands” that can look things up, calculate, and do things.

Turn a single answer into an observable, verifiable, and stoppable state machine

Understand fixed-length, structural, semantic, and parent-child chunking; handle overlap, context loss, tables, code, and embedding budgets.

From atomic claims, stable anchors, and entailment judgment to citation correctness, completeness, and source quality: build a truly verifiable answer.

From ontologies, entity resolution, and temporal edges to path retrieval and GraphRAG, understand when graphs outperform similarity and how they propagate errors.

Before answering, first retrieve relevant materials, letting the model 'answer with the book open' rather than making things up from memory.

Learn “to denoise from a blob of noise step by step” and “develop” the image.

Break business goals into observable attributes and locate failures caused by retrieval, generation, tools, and processes.

From sources and licensing, lineage, filtering, deduplication, and mixing ratios, to contamination, deletion requests, training snapshots, and incident response.

From confidence scores, reliability diagrams, and ECE to selective prediction, rejection thresholds, and distribution shift.

From five layers of control—input, context, output, tool, and runtime—to threshold cost, policy composition, and fault injection, understand what AI Guardrails can and cannot guarantee.

Hiding malicious instructions in content the model reads to hijack its behavior

From log-log straight lines and diminishing marginal returns to compute-optimal allocation, and then to data quality, capability thresholds, and total lifecycle cost.

Learn a new task using just a few examples in the prompt, without changing a single parameter

Comparing feature attribution, probes, counterfactuals, activation replacement, sparse features, and circuit analysis to understand fidelity, stability, completeness, and “readable does not equal true.”

From direct/indirect, multi-turn, encoding, suffix optimization, and multimodal jailbreaking to threat models, base rates, defense in depth, adaptive red teaming, and false rejection of legitimate requests.

Understand the capabilities, hyperparameters, and scale limitations of linear, RBF, and polynomial kernels through maximum margin, the dual problem, and the kernel trick.

From direct routing, small-then-large, and quality-gap prediction, to escalation thresholds, cascade latency, fairness risk, and online recalibration.

Understand vectors, distance metrics, HNSW/IVF/PQ, filtering, updates, and consistency, and separate index recall, latency, and end-to-end Retrieval-Augmented Generation (RAG).

Starting from the forward increment y=x+F(x) and the backward Jacobian I+JF, understand depth degradation, projection shortcuts, branch scaling, zero initialization, and Transformer's Pre-Norm/Post-Norm.

From continuous state equations, discretization, and convolution equivalence, to selective scanning, linear complexity, stability, and random-access boundaries.

From teacher demonstrations, programmatic ground truth, self-training, and augmentation, to verifier selection bias, coverage ratios, feedback loops, and model collapse.

Without changing the weights, use sampling, search, tools, and checking to convert extra computation into more reliable answers.

From video latent variables, spatiotemporal attention, and diffusion, to image-to-video, cascaded super-resolution, identity consistency, physics failures, long video planning, and temporal evaluation.

First distinguish the true distribution P from the model distribution Q, then derive cross-entropy, KL divergence, and perplexity step by step from self-information and entropy.

From system inventory, risk classification, and responsibility matrix to go-live gate, change approval, incident response, and decommissioning, understand how governance translates principles into lifecycle evidence.

Turn a sentence into a picture nobody has ever taken.

From TTFT/TPOT, arithmetic intensity, and continuous batching, to PagedAttention, FlashAttention, quantization, speculative decoding, and real-workload validation.

Let any tool and data plug into AI applications through the same standard interface.

从遮罩重绘、img2img噪声强度、注意力控制和指令编辑，到反演、局部性、身份一致、可逆性与编辑来源。

From mask inpainting, img2img noise strength, attention control, and instruction editing, to inversion, locality, identity consistency, reversibility, and edit provenance.

把注意力包装成可深度堆叠、可并行训练的标准积木

Packaging attention into a standard building block that can be stacked deeply and trained in parallel.

Learn from “input + target” samples and make reliable predictions on unseen data.

从「为什么需要」到「如何通过反向传播学会」

From “Why We Need It” to “How It Learns Through Backpropagation”