AirSOTA
Air School of Thoughts AtoZAirSOTA 知识矩阵：聚合大模型算法架构、科学育儿情境成长、加州地产考牌实战与全球数字化商业出海的权威专栏。
TalentMe · AI 学习与系统架构
工业级 AI 算法核心 69 题、前沿大模型系统架构演进与北美技术面试全流程备考深度长文。
高效微调 (PEFT) 架构全景：LoRA、QLoRA、DoRA、Prefix/Prompt Tuning、Adapters 与 MoRA/ReLoRA 深度剖析
> **核心摘要**：随着大语言模型 (LLM) 参数量迈向千亿级，全参数微调 (Full Fine-Tuning) 的显存与计算开销变得不可承受。**高效参数微调 (Parameter-Efficient Fine-Tuning, PEFT)** 技术通过仅冻结预训练基座权重 $W_0$，仅训练极少量的增量参数（通常
机器学习评估指标与数据工程全景：分类/回归/排序(NDCG)、概率校准、不平衡采样与文本预处理极客指南
> **核心摘要**：评估指标与数据预处理是连接模型输出与真实业务价值的数理基石。本指南全量整合 “ 中 9 大模块，涵盖分类评估（混淆矩阵、F-beta、Macro/Micro-F1、Hamming Loss）、搜索推荐排序评估（MRR、MAP、DCG/NDCG）、回归评估（MSE、RMSE、MAPE、$R^2$、
统计推断与假设检验：分布族、极大似然、中心极限定理、p-value、置信区间与功效分析全景
> **核心摘要**：统计推断是将带噪声的数据转化为经过校准的决策的学科，而假设检验是把证据转换为结论的引擎。本指南构建完整链路：常用分布族工具箱（正态、χ²、t、F、泊松、伯努利）；为大样本推断背书的**中心极限定理 (CLT)**；经由**极大似然估计 (MLE)** 与**矩估计**实现点估计并给出无偏性、一致性
经典与深度强化学习全景：MDP 体系、Bellman 最优方程、DQN、Policy Gradient、PPO 与 SAC 原理解构
> **核心摘要**：强化学习 (Reinforcement Learning, RL) 是研究智能体 (Agent) 在与动态环境交互过程中，如何通过试错 (Trial-and-Error) 学习最优策略 $pi(a|s)$ 以最大化累积折扣回报的数理科学。从经典表格型 Q-Learning 到结合深度神经网络的
搜索与计算广告系统设计:Query 意图理解、分布式倒排索引、RTB 竞价与 pCTR 预估
> **核心摘要**:搜索与广告是同一漏斗的两面:理解用户意图 → 海量候选检索 → 按相关性(广告还需按期望收入)排序 → 在严格延迟预算内混排与投放。搜索侧,本指南覆盖 Query 理解、分布式倒排索引、Recall@K / MRR / NDCG 指标族与 Pointwise/Pairwise/Listwise 学
DS Core Cheatsheet: Causal Inference, A/B Testing & Drift
> **Executive Summary**: The fundamental mandate of a Data Scientist (DS) is to drive business growth through rigorous statistical inference, causal identificat
RS Experiment Design & Reproducible Research: Ablations, Seed Control, Hyperparameter Search & the Full Reproducibility Checklist
> **Core Executive Summary**: In Research Scientist (RS) interviews — and in research itself — the difference between a credible paper and an irreproducible one
Tool Use & Function Calling: Toolformer Self-Taught Calls, JSON Schema & Sandbox Execution
> **Core Executive Summary**: LLMs cannot query real-time APIs or execute code natively. **Tool Use** and **Function Calling** bridge this gap by enabling LLMs
AI Safety & Privacy: Prompt Injection, Guardrails, Differential Privacy & Federated Learning
> **Core Executive Summary**: Interactive LLM applications introduce novel security threat vectors. **Prompt Injections** can hijack model behavior, and **PII l
Optimizers & Training Engineering Taxonomy: SGD, Momentum, AdamW Decoupled Weight Decay, Xavier/Kaiming Initialization & Gradient Checkpointing Guide
> **Summary**: Optimizers and training engineering bridge the gap between network architecture design and physical GPU memory limits. This 100% exhaustive guide
LLM Quantization & Model Compression: INT8/INT4 Mapping, SmoothQuant Outliers, GPTQ Hessian & AWQ/Distillation
> **Core Executive Summary**: As LLM parameter counts scale into tens to hundreds of billions, FP16/BF16 VRAM consumption and memory bandwidth become severe lat
Probabilistic Graphical Models: Naive Bayes, HMM Viterbi & Linear-Chain CRF Guide
> **Summary**: Probabilistic Graphical Models (PGM) combine graph theory and probability theory. This guide covers Naive Bayes conditional independence, HMM Vit
Speech & Audio Processing: Whisper Architecture, Log-Mel Spectrogram & Audio-LLM
> **Core Executive Summary**: Speech is the most natural medium for human interaction. Traditional Speech Recognition (ASR) relied on complex acoustic and langu
Model-Based RL & Planning: World Models, Dyna, MPC, MuZero & Dreamer
> **Core Executive Summary**: Model-based reinforcement learning (MBRL) equips the agent with an internal world model — a learned approximation of the transitio
TalentMe AI/ML/LLM Full Knowledge Taxonomy & Architecture Graph
> **Overview & Vision**: Modern Artificial Intelligence and Large Language Models have evolved into a massive, interdisciplinary, and mathematically rigorous en