Background knowledge

Knowledge for reading AI signals

Concepts that help you understand AI technology signals and their background.

Use this page when a signal refers to a concept you need to brush up on. Each concept says what it clarifies and which published signals depend on it.

27
concepts
6
foundation entries
22
linked skills
102
linked published signals

Knowledge entries are written in Chinese; signal titles are in English where the source is.

Foundation

Foundations

Start here when a signal uses an unfamiliar term or rests on an older idea.

Software architectureFoundation

API 契约与接口边界

当 AI 系统调用工具、串联服务或交换结构化上下文时,稳定的契约至关重要。

Helps you understandinterface boundaries, system trade-offs and integration risk.
Linked14 signals1 skills
Workflow
Open concept
DataFoundation

检索增强生成基础

为什么外部知识检索会改变答案的质量、时效性和信任边界。

Helps you understandassumptions about retrieval, freshness, trust and data flow.
Linked3 signals1 skills
Retrieval
Open concept
OperationsFoundation

人在回路的审查

智能体开始以机器速度产出之后,问题不再是「要不要人审」,而是「哪些决定值得占用一个人」——审查点放在哪一段,比有没有审查点更决定信任。

Helps you understandreview loops, observability and launch discipline.
Linked27 signals5 skills
Product strategyWorkflow
Open concept
Product thinkingFoundation

反馈闭环与团队学习

小而可见的闭环,比目标模糊的大版本发布更能让团队快速学习。

Helps you understandhow a technical change turns into scoped product and rollout decisions.
Linked9 signals3 skills
Product strategyObservability
Open concept
Product thinkingFoundation

副语言信号与语音交互

语气、停顿、情绪、说话人身份——这些转写文本里不存在的信息层,是语音交互与文字交互的本质差异,也是当前语音模型「会说不会听」差距的成因。

Helps you understandhow a technical change turns into scoped product and rollout decisions.
Linked4 signals1 skills
MultimodalProduct strategy
Open concept
Product thinkingFoundation

开放权重与模型许可

「开放权重」只说明模型文件可以下载,不说明能怎么用、训练数据是否公开——读一次开放发布,要分别看权重、许可和配套这三件事。

Helps you understandhow a technical change turns into scoped product and rollout decisions.
Linked17 signals4 skills
Open-weight models
Open concept

Intermediate

Intermediate patterns

Use these to compare approaches, trade-offs and implementation paths.

Software architectureIntermediate

工具使用与函数调用

许多新智能体系统背后的老思想:把推理与行动分开,并对两者分别验证——认出这个形状,才不会被协议和 SDK 的包装带走。

Helps you understandinterface boundaries, system trade-offs and integration risk.
Linked15 signals2 skills
AI agentsWorkflowAgent runtimes & protocols
Open concept
OperationsIntermediate

评估闭环

定义预期、观察实际、比较、收紧——评测的价值不在跑出一个数字,而在这个循环是不是真的在转。

Helps you understandreview loops, observability and launch discipline.
Linked28 signals4 skills
ObservabilityBenchmarks & evaluation
Open concept
Software architectureIntermediate

系统设计的权衡

当产品由 AI 驱动时,延迟、可靠性、成本和可控性这些权衡依然适用。

Helps you understandinterface boundaries, system trade-offs and integration risk.
Linked29 signals2 skills
WorkflowProduct strategy
Open concept
Machine learningIntermediate

模型选型与约束匹配

选模型不是挑最强的那个,而是先找出哪个约束最先咬人——延迟、隐私、成本,还是任务本身根本不需要那么大的模型。

Helps you understandmodel behaviour, the limits of evaluation, and practical constraints.
Linked22 signals6 skills
On-device AIProduct strategy
Open concept
Software architectureIntermediate

交互系统中的延迟权衡

交互系统里决定体验的不是平均延迟这一个数,而是用户在等待期间知道什么——语音与端侧 AI 让这条老经验重新变得要紧。

Helps you understandinterface boundaries, system trade-offs and integration risk.
Linked18 signals4 skills
On-device AIMultimodal
Open concept
Machine learningIntermediate

合成数据与数据配方

为什么新一代模型的能力越来越取决于「数据是怎么配出来的」,以及合成数据在其中扮演的角色与风险。

Helps you understandmodel behaviour, the limits of evaluation, and practical constraints.
Linked7 signals2 skills
Frontier modelsAI agentsTraining & post-training
Open concept
Machine learningIntermediate

对抗性评估与红队方法

一种源自安全工程的经典思路:不测系统平均表现,而是主动寻找让它失败的输入——理解现代模型安全测试与注入防御的方法论根源。

Helps you understandmodel behaviour, the limits of evaluation, and practical constraints.
Linked15 signals2 skills
AI agentsFrontier modelsSafety & alignment+1
Open concept
Machine learningIntermediate

混合专家模型(MoE)架构

理解稀疏激活如何让总参数、推理算力与显存占用三者脱钩——读懂开源旗舰模型规格表和估算部署成本的前提概念。

Helps you understandmodel behaviour, the limits of evaluation, and practical constraints.
Linked12 signals5 skills
Frontier modelsInference & deployment
Open concept
Machine learningIntermediate

模型量化与数值精度

把权重和激活从高精度压到低比特,是让大模型跑进有限显存的主要手段;代价是精度损失,而损失落在哪一层决定了它是无感还是灾难。

Helps you understandmodel behaviour, the limits of evaluation, and practical constraints.
Linked13 signals5 skills
Inference & deploymentOn-device AIQuantisation & speculative decoding
Open concept
Machine learningIntermediate

推测解码与推理加速技术

自回归生成的逐 token 串行是大模型推理延迟的根源;推测解码、连续批处理与 KV 缓存管理从算法与系统两层,把同样的硬件榨出数倍吞吐。

Helps you understandmodel behaviour, the limits of evaluation, and practical constraints.
Linked14 signals4 skills
Inference & deploymentFrontier modelsQuantisation & speculative decoding
Open concept
OperationsIntermediate

完成判定与验收信号

长时程智能体最难回答的不是「这一步怎么做」,而是「这一步做完了没有」。完成信号从哪里来、由谁定义、能不能被验证,决定了它能不能被托付真正的工作。

Helps you understandreview loops, observability and launch discipline.
Linked30 signals5 skills
ObservabilityAI agentsBenchmarks & evaluation
Open concept
Software architectureIntermediate

本地与云混合推理架构

把「哪些请求留在本地、哪些上云」当成一个架构决策:隐私、延迟、成本与能力四个约束的显式权衡。

Helps you understandinterface boundaries, system trade-offs and integration risk.
Linked11 signals4 skills
On-device AIInference & deployment
Open concept
Software architectureIntermediate

长时程智能体的记忆与上下文管理

上下文窗口不是记忆。长时程任务需要在有限窗口之上构建工作记忆、外部存储与检索回填的分层体系,其设计质量直接决定智能体能走多远。

Helps you understandinterface boundaries, system trade-offs and integration risk.
Linked14 signals2 skills
AI agentsWorkflow
Open concept
Machine learningIntermediate

知识蒸馏

用大模型(教师)的输出去训练小模型(学生),让小模型以低得多的成本接近大模型的表现——多数小型开放模型都是这样来的。

Helps you understandmodel behaviour, the limits of evaluation, and practical constraints.
Linked6 signals3 skills
Training & post-trainingOpen-weight models
Open concept
Machine learningIntermediate

监督微调与参数高效微调(LoRA)

在训练好的模型上用一批示范数据继续训练,让它适应特定任务或风格;LoRA 只训练一小组附加参数,让个人和小团队也做得起。

Helps you understandmodel behaviour, the limits of evaluation, and practical constraints.
Linked11 signals2 skills
Training & post-trainingOpen-weight models
Open concept
Software architectureIntermediate

智能体沙箱与执行隔离

智能体会执行它自己写的代码和命令,沙箱决定这些动作最多能波及哪里——隔离强度、网络出口和凭据范围是三个要分开看的维度。

Helps you understandinterface boundaries, system trade-offs and integration risk.
Linked15 signals5 skills
Agent runtimes & protocolsCoding agentsSafety & alignment
Open concept
OperationsIntermediate

基于执行的代码评测

让模型在真实代码仓库或终端里完成任务,再用测试判定对错——SWE-bench、Terminal-Bench 这类分数说明了什么、哪些能比、哪些不能比。

Helps you understandreview loops, observability and launch discipline.
Linked11 signals5 skills
Coding agentsBenchmarks & evaluation
Open concept
OperationsIntermediate

可复现性与可验证结果

一个结果要能被别人用同样的材料重做出来才站得住——AI 做科研、AI 评测 AI 的结论越来越多,判断它们可不可信,先看这一点。

Helps you understandreview loops, observability and launch discipline.
Linked9 signals4 skills
AI for scienceBenchmarks & evaluation
Open concept

Advanced

Advanced background

Use these when a signal changes assumptions about architecture, data or technical strategy.

DataAdvanced

面向知识系统的图思维

把概念和关联显式建模,能让某些推理更易于检查和维护。

Helps you understandassumptions about retrieval, freshness, trust and data flow.
Linked4 signals1 skills
Knowledge graphsRetrieval
Open concept
Machine learningAdvanced

KV 缓存与长上下文推理

生成每个新 token 都要回看前面所有 token;KV 缓存把这份回看要用的中间结果存下来,于是上下文越长,显存和带宽越紧——读懂推理引擎更新和长上下文定价的前提。

Helps you understandmodel behaviour, the limits of evaluation, and practical constraints.
Linked12 signals4 skills
Quantisation & speculative decodingInference & deployment
Open concept
Machine learningAdvanced

强化学习后训练与可验证奖励

让模型自己作答、按结果打分、再朝高分方向更新——近一年推理与智能体能力的主要来源,也是奖励投机这类意外行为的源头。

Helps you understandmodel behaviour, the limits of evaluation, and practical constraints.
Linked12 signals2 skills
Training & post-trainingFrontier models
Open concept