01Variational Autoencoders02Generative Adversarial Networks03Normalizing Flows04Diffusion Models05Autoregressive Generative Models06Image & Video Generation073D & Multimodal Generation08Multimodal Foundation Models
XIAI Agents & Autonomous Systems
01Agent Fundamentals02LLM-Based Agents03Tool Use & Function Calling04Memory & Knowledge Management05Planning & Reasoning06Multi-Agent Systems07Agent Evaluation & Benchmarking
XIIRobotics & Embodied AI
01Robot Perception & Sensing02Motion Planning & Control03Learning from Demonstration & Imitation04Sim-to-Real Transfer05Foundation Models for Robotics06Autonomous Vehicles
XIIISpecialized ML Methods
01Time Series Analysis & Forecasting02Anomaly Detection03Causal Inference04Causal Machine Learning05Graph Neural Networks06Survival Analysis & Event Modeling07Bayesian Deep Learning08Meta-Learning & Few-Shot Learning09Continual & Lifelong Learning10Federated Learning & Privacy-Preserving ML11Neurosymbolic AI
XIVApplied Domains
01Recommender Systems02Search & Information Retrieval03Financial ML & Quantitative Methods04Healthcare & Clinical AI05AI for Cybersecurity06AI for Education & Personalization07AI for Manufacturing & Operations08Human-AI Interaction & UX
XVAI for Science
01Scientific Machine Learning02AI for Biology & Genomics03AI for Drug Discovery & Molecular Design04AI for Protein Science05AI for Climate & Earth Systems06AI for Physics, Materials & Astronomy
XVIMLOps & Production ML
01Experiment Tracking & Reproducibility02Feature Stores & Data Management for ML03Model Deployment & Serving04Model Monitoring & Drift Detection05CI/CD for Machine Learning06A/B Testing & Causal Experimentation07Responsible Release & Deployment Practices
XVIIAI Infrastructure & Systems
01Hardware for ML02Distributed Training03Model Compression04Inference Optimization05AI Chips & Custom Silicon
XVIIIAI Safety, Alignment & Governance
01AI Safety Fundamentals02Technical Alignment Methods03Robustness & Adversarial ML04Mechanistic Interpretability05Explainability for Practitioners06Fairness, Bias & Equity07Privacy in ML08AI Governance, Policy & Regulation
A Reading Guide · 2026 Edition
How to actually learn AI & data science — without drowning in papers.
Most of the historic papers in this field are better read about than read. A well-written chapter can compress a paper's core idea, place it in context, and spare you the outdated notation. This guide picks one or two strong books for each topic, notes the handful of papers still worth reading in the original, and points to the blogs, courses, and technical reports that fill in wherever the books run out of road.
How to use this guide
Each of the nine chapters below is built around one primary book that you'd read end-to-end if you were serious about the topic, plus secondary books for alternate angles or deeper dives. Below the books you'll find free online resources (courses, blogs, docs) that are genuinely excellent, then a short list of original papers still worth reading — the ones where the paper is clearer, funnier, or more surprising than any textbook treatment. Each chapter closes with modern extras: work so recent that no book covers it yet.
The color key: primary book is amber, secondary books are teal, practical / applied books are blue, and "book-ish long-form" resources are violet. Anything marked free is legitimately free online — a PDF released by the publisher, author website, or project documentation.
Chapter One
Foundations of AI — philosophy, history, first principles
Most of the field's early papers were written to stake out territory nobody had thought to occupy yet. Read about them before you read them: a good historical chapter will give you the map, and then the one or two papers that still reward a direct visit become much shorter than you thought.
The main book
Artificial Intelligence: A Modern Approach
Stuart Russell & Peter Norvig·4th ed. · 2020·Pearson
The undergraduate AI textbook, now in its fourth edition. Encyclopedic in scope: search, logic, planning, probabilistic reasoning, machine learning, natural language, vision, robotics, and philosophy. Decades of pedagogical refinement have made it the most reliable single starting point in the field.
Covers (in book form): Turing's imitation game and its philosophical descendants; the Dartmouth workshop and symbolic AI; Samuel's checkers and the origins of machine learning; the Newell–Simon physical-symbol-system hypothesis; the Minsky–Papert perceptron critique; Quinlan's decision trees. Every foundational paper in this chapter is digested somewhere in the first twelve chapters or the closing chapters on philosophy and ethics.
Why Machines Learn: The Elegant Math Behind Modern AI
Anil Ananthaswamy·2024·Dutton
Popular but mathematically serious. Walks from perceptrons to transformers explaining the ideas in their historical order. The best single book for a bright non-specialist, and a useful palate-cleanser even for experts.
The Deep Learning Revolution
Terrence Sejnowski·2018·MIT Press
Insider's memoir of the neural-network tradition from the inside: Hinton's basement, Hopfield nets, Boltzmann machines, the slow warming of the long winter. Light on math, heavy on the culture and personalities that shaped modern AI.
Free online resources
MIT 6.034 — Artificial IntelligenceOpenCourseWare · Patrick Winston
Winston's lectures are still the best entry-level AI course on the internet. Why: Winston tells the story of symbolic AI with the affection of someone who lived it.
Modern companion to AIMA with its own superb Pac-Man-based assignments.
"The Bitter Lesson"Rich Sutton · 2019 · essay
One page. More influential than most textbook chapters. Why: it names a pattern — general methods that leverage computation beat handcrafted ones — that every subsequent chapter of this guide illustrates.
Original papers still worth reading
Computing Machinery and IntelligenceTuring · 1950
Why in the original: short, witty, weirdly prescient about the objections future critics would raise. Every textbook paraphrase loses the voice.
A Mathematical Theory of CommunicationShannon · 1948
Why in the original: if you've ever wondered where cross-entropy came from or why we measure things in bits, reading the first few sections demystifies a lot at once. Pair it with James Gleick's The Information as an entertaining companion.
Chapter Two
Classical Machine Learning — statistics, trees, and tabular data
The quiet giant. For most applied problems outside frontier AI, well-tuned gradient boosting still beats neural networks, and linear models still beat gradient boosting when interpretation matters. The original papers here are almost all covered more clearly in one of two widely-loved textbooks.
The main book
An Introduction to Statistical Learning with Applications in Python
Gareth James, Daniela Witten, Trevor Hastie & Robert Tibshirani·2023·Springer
The book to read first. Covers linear and logistic regression, resampling and model selection, trees, boosting and random forests, SVMs, unsupervised learning, and a gentle introduction to deep learning — all at a pace a careful reader can actually finish. There's an R edition (2013/2021) and this Python edition.
Covers (in book form): Fisher's LDA, Lloyd's k-means, Breiman's CART, Cortes–Vapnik SVMs, Freund–Schapire AdaBoost, Breiman's Random Forests, Friedman's gradient boosting.
Trevor Hastie, Robert Tibshirani & Jerome Friedman·2nd ed. · 2009·Springer
The rigorous older sibling of ISL. Same authors, more math, more depth, more breadth. The definitive reference for classical ML; the chapter on boosting is essentially the Friedman paper rewritten for humans.
Probabilistic Machine Learning: An Introduction & Advanced Topics
Kevin P. Murphy·2022 & 2023·MIT Press
Two-volume modern reference, successor to Murphy's 2012 book. Vol 1 covers foundations through deep learning; Vol 2 covers Bayesian inference, causality, generative models, and decision making. If you want one bookshelf that spans from Fisher to diffusion models, this is it.
Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow
Aurélien Géron·3rd ed. · 2022·O'Reilly
The best applied book for this chapter. Read ISL for the ideas, read Géron alongside to actually implement them. Covers XGBoost, cross-validation recipes, and the engineering around making classical methods work in practice.
Free online resources
scikit-learn User Guidescikit-learn.org
A textbook in its own right, with working code for every method. Read the "User Guide" tab, not just the API reference.
Kaggle Learnkaggle.com/learn
Short, interactive, code-first courses. The best way to turn book knowledge into muscle memory.
XGBoost documentation — "Introduction to Boosted Trees"xgboost.readthedocs.io
Tianqi Chen's tutorial page is the cleanest single explanation of gradient boosting you'll find.
Original papers still worth reading
Statistical Modeling: The Two CulturesBreiman · 2001
Why in the original: it's an essay, not a technique, and no book summary replaces the voice. Required reading for anyone straddling statistics and ML.
XGBoost: A Scalable Tree Boosting SystemChen & Guestrin · 2016
Why in the original: unusually well-written systems paper. The engineering details (cache-aware access, sparsity handling, out-of-core) are why XGBoost won so many competitions, and textbooks skip them.
Modern extras not yet in books
Tabular Data: Deep Learning is Not All You NeedShwartz-Ziv & Armon · 2021 · arXiv
Empirical receipt for why gradient boosting keeps winning on tabular data.
TabPFNHollmann et al. · 2023–2025
A surprisingly strong pretrained transformer for small-data tabular classification. Worth skimming as a glimpse of where the field may be going.
Chapter Three
Deep Learning Architectures — networks, optimizers, training
A remarkable share of "classic" deep learning papers — LeNet, AlexNet, ResNet, backprop, LSTM, dropout, Adam, batch norm — are better experienced as chapters in a modern textbook than as the original publications. Their ideas are stable; their notation and framing aren't.
The main book
Understanding Deep Learning
Simon J. D. Prince·2023·MIT Press
The best modern deep-learning textbook. Beautifully typeset, richly illustrated, careful about intuition and mathematics in equal measure. Covers MLPs, CNNs, RNNs, transformers, graph networks, reinforcement learning, generative models (GANs, VAEs, normalizing flows, diffusion). Replaces a small stack of older books.
Covers (in book form): the perceptron and backpropagation, LSTMs, LeNet→AlexNet→ResNet, dropout and batch norm, the Adam optimizer. Every paper in the "Deep Learning Architectures" category of the papers guide appears somewhere in its pages.
Aston Zhang, Zachary Lipton, Mu Li & Alexander Smola·continuously updated·Cambridge / d2l.ai
Interactive textbook with every example available in PyTorch, MXNet, TensorFlow, and JAX. The rare book maintained like software: updates for new architectures land within months. A good "code first" counterpart to Prince's more conceptual treatment.
Ian Goodfellow, Yoshua Bengio & Aaron Courville·2016·MIT Press
The "Goodfellow book" — still great on mathematical foundations (linear algebra, probability, information theory, numerical methods). Its architecture chapters are dated (pre-transformer), so treat it as a foundations reference rather than a current-state survey.
Neural Networks: Zero to HeroAndrej Karpathy · YouTube
Builds a small GPT from scratch in Python, live. If you watch one video series in your life about deep learning, make it this one. Why: Karpathy teaches the way an engineer learns — from code outward.
Practical Deep Learning for Codersfast.ai · Jeremy Howard & Sylvain Gugger
Top-down, code-first. Famously gets students shipping models in week one.
Stanford CS231n & Michigan EECS 498Justin Johnson
Johnson's two courses are the definitive academic treatment of deep learning applied to vision. The Michigan version is the newer, cleaner one.
Original papers still worth reading
Deep Residual Learning for Image RecognitionHe, Zhang, Ren & Sun · 2015
Why in the original: nine pages explain an idea so clean (add a skip connection) that rewriting it in a textbook always makes it sound more complicated than it is.
Batch NormalizationIoffe & Szegedy · 2015
Why in the original: the "internal covariate shift" framing turned out to be wrong, but the empirical tables are still the best argument for the method. A good paper to read critically.
Modern extras not yet in books
FlashAttention & FlashAttention-2 / 3Tri Dao · 2022, 2023, 2024
The systems paper series that made long-context transformers tractable. Still missing from most textbooks.
The Annotated TransformerSasha Rush & Austin Huang · HarvardNLP
The "Attention Is All You Need" paper rebuilt as runnable code with commentary. The gold-standard implementation walkthrough.
Chapter Four
Natural Language & Transformers — embeddings to LLMs
The fastest-moving chapter in this guide. The core ideas have books now, but the production-relevant details — retrieval, fine-tuning, agents, evals — mostly live in blogs and courses. Read the book to understand how an LLM works; read everything else to understand how to use one.
The main book
Hands-On Large Language Models
Jay Alammar & Maarten Grootendorst·2024·O'Reilly
Alammar is the author of the famous "Illustrated Transformer" blog posts, and this book extends that visual, intuition-first style into a coherent end-to-end treatment. Covers tokenization, attention, prompting, retrieval, fine-tuning, RLHF, agents, and evals — the whole modern stack.
Covers (in book form): the Transformer itself, BERT-style pretraining, GPT-3 few-shot prompting, InstructGPT-style RLHF, LoRA fine-tuning.
Strong companions
Build a Large Language Model (From Scratch)
Sebastian Raschka·2024·Manning
Implements a GPT-2 sized model in PyTorch from tokenization to generation to instruction fine-tuning, with clear code. The pedagogical complement to Alammar & Grootendorst: one explains, the other constructs.
Speech and Language Processing
Dan Jurafsky & James H. Martin·3rd ed. (draft)·Stanford
The canonical NLP textbook, updated continuously as free draft chapters. Covers both traditional NLP (parsing, tagging, speech) and modern transformer-based methods. Authoritative and thorough in a way that no industry book can afford to be.
Not strictly an "NLP" book, but the best single reference for designing products on top of LLMs: evaluation, prompt engineering, RAG, fine-tuning economics, inference optimization, and post-deployment monitoring.
Free online resources
Stanford CS336 — Language Modeling from ScratchTatsunori Hashimoto & Percy Liang · 2024+
The best current academic course on LLMs. Builds one end-to-end. Lectures and assignments are public.
Jay Alammar's "Illustrated" seriesjalammar.github.io
The Illustrated Transformer, Illustrated GPT-2, Illustrated BERT, Illustrated Stable Diffusion. Still the clearest visual explanations on the internet.
Hugging Face NLP & LLM courseshuggingface.co/learn
Free, hands-on, updated. Especially strong on the practical "how do I fine-tune this" layer.
Lilian Weng's bloglilianweng.github.io
Long, careful, textbook-quality posts on attention variants, prompt engineering, agents, and more. Cited in papers.
Original papers still worth reading
Attention Is All You NeedVaswani et al. · 2017
Why in the original: ten pages, clearer than most textbook treatments. The hardest part of the modern stack to understand is easiest to learn from its own paper.
Scaling Laws for Neural Language ModelsKaplan et al. · 2020
Why in the original: the specific plots are the argument. Books tend to summarize the conclusion and skip the evidence.
Training Compute-Optimal Large Language ModelsHoffmann et al. ("Chinchilla") · 2022
Why in the original: revises Kaplan in a way most books haven't caught up to. If you read one, read both.
The architecture decision underneath many frontier models. Most textbooks still treat dense transformers as the default.
The reasoning-model papers & system cardsOpenAI o1 (2024), DeepSeek-R1 (2025), etc.
The shift to "test-time compute" is too recent for any book. The system cards themselves are the primary source.
"Building Effective Agents"Anthropic engineering blog · 2024
Short, practical taxonomy of agent patterns from a frontier lab. Already cited like a textbook.
Chapter Five
Computer Vision — classical to multimodal
The most textbook-mature sub-field of deep learning, because vision benefited from thirty years of classical research before the deep-learning wave. One big book covers almost everything; a couple of courses cover the rest.
The main book
Computer Vision: Algorithms and Applications
Richard Szeliski·2nd ed. · 2022·Springer
Comprehensive is the word. Covers image formation, classical features, optical flow, structure from motion, stereo, segmentation, recognition, and a solid deep-learning chapter brought up to the late-2010s. Written by a veteran of Microsoft Research with unusual patience for historical context.
Covers (in book form): Lucas–Kanade optical flow, Lowe's SIFT, VGG, GoogLeNet, U-Net, YOLO, Mask R-CNN, Vision Transformers. Essentially the whole computer-vision category of the papers guide.
Less encyclopedic, more hands-on. Good for readers who want to actually train detectors and classifiers rather than survey the field.
Multiple View Geometry in Computer Vision
Richard Hartley & Andrew Zisserman·2nd ed. · 2003·Cambridge
The definitive classical reference for 3D geometry — epipolar lines, bundle adjustment, structure from motion. Old but unreplaced. Essential if you care about SLAM, photogrammetry, or modern 3D reconstruction.
Free online resources
Stanford CS231n & Michigan EECS 498Justin Johnson
The gold-standard deep-learning-for-vision courses. EECS 498 is the newer recording; CS231n is the foundational version taught by Andrej Karpathy and Fei-Fei Li.
PyImageSearch blogAdrian Rosebrock
The practitioner's go-to for running modern CV pipelines end-to-end, with complete code.
The Roboflow blogroboflow.com/blog
Up-to-date hands-on tutorials for current detection, segmentation, and VLM pipelines.
Original papers still worth reading
U-NetRonneberger, Fischer & Brox · 2015
Why in the original: eight pages, beautifully clear. Every diffusion model you use has a U-Net at its core, and this is where to see the architecture explained straight.
Learning Transferable Visual Models From Natural Language Supervision (CLIP)Radford et al. · 2021
Why in the original: the zero-shot evaluation tables and the training-data discussion are the paper's most-cited pieces, and books rarely reproduce them in full.
Modern extras not yet in books
Segment Anything (SAM) and SAM 2Kirillov et al. · 2023, 2024 · Meta
A foundation model for segmentation that reset the field. SAM 2 extends it to video.
3D Gaussian Splatting for Real-Time Radiance Field RenderingKerbl et al. · 2023
The technique that quietly replaced NeRF for most practical 3D-reconstruction tasks.
DINOv2 / DINOv3Meta AI · 2023, 2024
The best open self-supervised vision encoders; the default choice for downstream tasks as of this writing.
Chapter Six
Reinforcement Learning — MDPs to RLHF
The field has one textbook so good that almost no one writes a second — and then a small library of courses and blog posts for everything deep RL has added on top. RL has re-entered the mainstream through LLM fine-tuning, so the reading list now stretches from tabular Q-learning to GRPO.
The main book
Reinforcement Learning: An Introduction
Richard S. Sutton & Andrew G. Barto·2nd ed. · 2018·MIT Press
The canonical text. Sutton and Barto invented much of what they're writing about. Covers bandits, MDPs, dynamic programming, Monte Carlo, temporal-difference learning, Q-learning, on- and off-policy methods, policy gradients, and planning. Paced for reading, not reference.
Covers (in book form): Sutton's TD learning, Watkins' Q-learning, Tesauro's TD-Gammon, the whole pre-deep-RL lineage.
Mykel J. Kochenderfer, Tim A. Wheeler & Kyle H. Wray·2022·MIT Press
Broader than Sutton–Barto: MDPs, POMDPs, multi-agent, exploration under uncertainty. Beautifully typeset, with Julia code. Good second book, especially for anyone applying RL to real decision problems rather than games.
The best applied book. Walks through DQN, policy gradients, PPO, and RLHF with working PyTorch code. The third edition adds modern topics including LLM fine-tuning.
Free online resources
David Silver — UCL RL Course10 lectures · YouTube
Silver (co-founder of DeepMind, architect of AlphaGo) walks through Sutton-Barto on a whiteboard. Still the best starter course two decades on.
OpenAI Spinning Up in Deep RLspinningup.openai.com
A structured reading list and implementation guide for the main deep-RL algorithms, written to a very high standard.
Berkeley CS285 — Deep Reinforcement LearningSergey Levine
The graduate course. Heavier on theory and robotics applications than OpenAI's guide.
Original papers still worth reading
Playing Atari with Deep Reinforcement Learning (DQN)Mnih et al. · 2013
Why in the original: nine pages; the pseudo-code is clear; the figure of DQN learning Breakout is iconic.
Proximal Policy Optimization AlgorithmsSchulman et al. · 2017
Why in the original: short, practical, and still a reference that working RL engineers re-read. Books cover it briefly; the paper has the hyperparameter discussion you'll actually need.
Deep Reinforcement Learning from Human PreferencesChristiano et al. · 2017
Why in the original: the foundational RLHF paper. Not yet in most books and directly relevant to anyone working on language models.
Modern extras not yet in books
DeepSeek-R1 & GRPODeepSeek · 2025
The RL algorithm family now used to train reasoning LLMs. Replaces PPO in much of the post-training pipeline.
"The 37 Implementation Details of Proximal Policy Optimization"ICLR Blog Post · Huang et al. · 2022
The best single document on why PPO implementations differ from the paper — and why yours probably doesn't work yet.
Chapter Seven
Generative Models — VAEs, GANs, diffusion, and what comes next
Of all the chapters in this guide, this is the one where the field moves fastest and where blog posts sometimes explain the math better than papers. A single modern book plus Lilian Weng's archive will get you most of the way.
The main book
Generative Deep Learning
David Foster·2nd ed. · 2023·O'Reilly
The best single book on modern generative modeling. Covers autoencoders, VAEs, GANs (DCGAN through StyleGAN), autoregressive models, normalizing flows, energy-based models, diffusion, transformer-based generation, music generation, and world models. Working Keras code throughout.
Covers (in book form): the VAE, the original GAN and DCGAN, StyleGAN, DDPM, conditional diffusion, and latent diffusion (Stable Diffusion).
Newer companion
Hands-On Generative AI with Transformers and Diffusion Models
Pedro Cuenca, Apolinário Passos, Omar Sanseviero & Jonathan Whitaker·2024·O'Reilly
Written by the Hugging Face team. More practitioner-oriented than Foster, and more up-to-date on the diffusion side (rectified flow, SDXL-era models, ControlNet, fine-tuning).
Free online resources
Hugging Face Diffusion Models Coursehuggingface.co/learn
Free, code-first, from DDPM up through latent diffusion. The most efficient way to understand diffusion by implementing it.
Lilian Weng — "What are Diffusion Models?" and "From GAN to WGAN"lilianweng.github.io
Two of the most-cited blog posts in the field. Why: clearer than the papers they summarize.
Yang Song — "Generative Modeling by Estimating Gradients of the Data Distribution"yang-song.net/blog
The score-based view of diffusion, explained by one of the people who invented it.
Why in the original: the Bayesian derivation is genuinely elegant, and most book treatments skip it in favor of "just train a denoiser." Worth the effort.
Generative Adversarial NetsGoodfellow et al. · 2014
Why in the original: an eight-page classic; history is worth a half hour.
The training objective now used by Stable Diffusion 3, Flux, and many video models. Books haven't caught up.
Sora / Veo / Movie Gen technical reportsOpenAI 2024 · Google 2024 · Meta 2024
The current state of the art in video generation. The reports, not the glossy demos, are where the ideas are.
Consistency Models & LCMSong et al. 2023, Luo et al. 2023
The set of techniques that made real-time diffusion sampling possible.
Chapter Eight
Data Systems & Infrastructure — data engineering and MLOps
Every model in every other chapter of this guide runs on top of the machinery in this one. The canonical Google systems papers (GFS, MapReduce, Bigtable, Dynamo, Spark) are best read through a modern synthesis rather than one at a time — someone has written the textbook so that you don't have to.
The main book
Designing Data-Intensive Applications
Martin Kleppmann·2017 (2nd ed. in progress)·O'Reilly
The reference every working data engineer cites. Kleppmann weaves the canonical systems papers into a single coherent narrative about reliability, scalability, storage, consistency, and stream processing. A second edition is being released chapter-by-chapter through the O'Reilly early-access program.
Covers (in book form): GFS and distributed file systems, MapReduce and batch processing, Bigtable and column stores, Dynamo and the CAP trade-offs, Spark RDDs and stream processing. The "systems" half of the data chapter of the papers guide is effectively this book.
Data engineering & ML systems
Fundamentals of Data Engineering
Joe Reis & Matt Housley·2022·O'Reilly
Complement to Kleppmann: less about database internals, more about the modern data stack — ingestion, transformation, warehousing, orchestration, and governance. The best single book for someone entering the data-engineering profession today.
Designing Machine Learning Systems
Chip Huyen·2022·O'Reilly
The MLOps textbook. Effectively, this is the "Hidden Technical Debt" paper grown into a whole book: data pipelines, feature stores, training/serving skew, monitoring, drift, and the organizational patterns that actually ship models.
AI Engineering
Chip Huyen·2024·O'Reilly
The LLM-era companion to Designing ML Systems. Covers prompt management, evaluation frameworks, retrieval, inference economics, and deployment — the infrastructure concerns specific to shipping products built on foundation models.
Free online resources
madewithml.comGoku Mohandas
A free, hands-on course covering the full production ML lifecycle, including CI/CD for models and API design.
dbt Learngetdbt.com/learn
If your work touches analytics or modern data stacks, this is essentially the industry-standard onboarding.
Apache Arrow documentation & Voltron Data blogarrow.apache.org
The invisible substrate of modern analytics; the Arrow columnar format is worth understanding at a glance.
Original papers still worth reading
Hidden Technical Debt in Machine Learning SystemsSculley et al. · 2015
Why in the original: short, pointed, still the clearest enumeration of ML failure modes in production. Worth re-reading every year.
MapReduce: Simplified Data Processing on Large ClustersDean & Ghemawat · 2004
Why in the original: historically important, remarkably readable, and shorter than its Kleppmann chapter.
Modern extras not yet in books
"Building LLM Applications for Production"Chip Huyen · 2023 · blog post
Long-form blog post that became a reference document for LLM product engineers.
vLLM & PagedAttentionKwon et al. · 2023
The serving system behind most open-weight LLM deployments; the paper is also a well-written explanation of inference-time memory management.
Chapter Nine
AI Safety, Alignment & Ethics — the harder questions
This is the chapter where books are weakest, because the field is young and changes as models change. The right approach is one accessible narrative book to get your bearings, one technical textbook for the research agenda, and an ongoing subscription to a few research blogs and forums.
The main book
The Alignment Problem
Brian Christian·2020·W. W. Norton
The most accessible narrative of modern ML safety and alignment. Christian interviews dozens of researchers and turns their work into a readable story covering fairness, robustness, reward hacking, and value alignment. Starts with Word2Vec gender bias and ends at RLHF.
The technical textbook
Introduction to AI Safety, Ethics, and Society
Dan Hendrycks·2024·Center for AI Safety
The first real textbook of AI safety as an academic field. Covers risk analysis, alignment techniques, governance, catastrophic-risk scenarios, and the empirical literature. Free online, actively maintained.
Human Compatible: Artificial Intelligence and the Problem of Control
Stuart Russell·2019·Viking
Russell (co-author of AIMA) argues that classical AI's fixed-objective paradigm is unsafe and proposes a research program of "provably beneficial" AI built around uncertainty over human preferences. Essential framing, even where you disagree.
Atlas of AI
Kate Crawford·2021·Yale
The political-economy view: AI as an extractive industry built on data, labor, energy, and state power. A sharp counterweight to the internalist technical literature.
Weapons of Math Destruction
Cathy O'Neil·2016·Crown
Slightly dated but still the most accessible introduction to algorithmic harm. Essential if your work touches consumer-facing ML.
Free online resources
AI Alignment Forumalignmentforum.org
The main long-form venue for technical alignment research outside of industry labs. High-signal, high-density.
Anthropic research bloganthropic.com/research
One of the two or three main sources of frontier-lab safety work. The "Core Views on AI Safety" post is the best single summary of the lab's agenda.
The canonical body of work on mechanistic interpretability, written as a public research diary rather than a series of papers.
Neel Nanda — "A Comprehensive Mechanistic Interpretability Guide"neelnanda.io
The best self-study entry point for mech interp.
Original papers still worth reading
Concrete Problems in AI SafetyAmodei, Olah, Steinhardt, Christiano, Schulman & Mané · 2016
Why in the original: a research-agenda paper that defined an entire subfield. Reads like a field guide; every subsequent safety paper is in conversation with it.
On the Dangers of Stochastic ParrotsBender, Gebru, McMillan-Major & Shmitchell · 2021
Why in the original: the rhetoric is the argument. Reading the authors' own sentences is a different experience from reading someone else's summary.
Modern extras not yet in books
Responsible Scaling PoliciesAnthropic, OpenAI, Google DeepMind · 2023+
Labs' published commitments on capability thresholds and safety testing. The governance conversation now orbits these documents.
Sleeper AgentsHubinger et al. (Anthropic) · 2024
An empirical study of models that behave deceptively under training — one of the papers that reframed the "alignment faking" conversation.
Scaling Monosemanticity & Circuit TracingAnthropic interpretability team · 2024, 2025
The state of the art in interpreting a real production LLM. Best read as a sequence, in the order the posts appeared.
Richard Ngo — "AGI Safety from First Principles"LessWrong sequence · 2020
Still the clearest conceptual argument for why capable AI systems might be catastrophically difficult to align. Long, but careful.
A note on reading order
If you are starting from scratch and plan to read most of this guide, a reasonable sequence is:
An Introduction to Statistical Learning first, to get comfortable with the vocabulary of models and evaluation; then Understanding Deep Learning, which supersedes a decade of older books; then the two LLM books (Hands-On Large Language Models and Build a Large Language Model from Scratch) in parallel, one for ideas and one for code. Add Reinforcement Learning: An Introduction when you want to understand how modern models are fine-tuned, and Designing Data-Intensive Applications when you want to put them into production. The Alignment Problem and Introduction to AI Safety can be read at any point — they get more interesting the more of the rest you've already absorbed.
Last revised: April 2026. A companion document to the landmark-papers guide and the landscape map.