Alignment, Preferences, and Oversight
For teams training assistants from preference data and evaluating whether alignment survives real deployment pressure.
Part of Language AI Handbook. The online chapters are free to read; the PDF edition is yours to keep, with future updates included.
Become a more confident language AI practitioner.
Build the understanding to explain how language models work and make more informed choices about using them. The handbook connects the foundations to the methods behind modern language AI, with explanations you can return to as you learn.
See what you’ll learnWhat you will learn
- Understand Alignment and RLHF.
- Understand Scalable Oversight and Safety Training.
Who this volume is for
- Engineers, researchers, students, AI enthusiasts, linguists, product managers, and anyone interested in understanding or building modern language AI systems, from foundational NLP to advanced large language models.
Prerequisites: Assumes Volume 11 and basic reinforcement learning concepts.
Volume 12 PDF edition
Own this 510-page edition
Keep the PDF for offline reading, with future updates included. The corresponding chapters stay free online.
- Carefully typeset PDF
- Future updates and errata included
- Online chapters remain free to read
Volume 12 of 20 · In progress
$24
one-timeVersion 2026.08.0 · 510 pages · secure checkout via Stripe
Delivered by email · Free updates included
Want the complete library?
Access to every volume, delivered as separate PDFs. Save $257 vs. buying available volumes separately.
Get all-volume access · $199Reader verdict
5.0
out of 5
12 five-star reader reviews
Genuine feedback, with names abbreviated for privacy.
Meet the community“Really appreciate the way you have explained the concepts. Simple to understand, yet builds up knowledge exponentially, all while reinforcing it with examples repeatedly. I've read the last two chapters `74.Self-attention & 75.Q,K,V` and I've to admit, my concepts have never been clearer. You really are doing an amazing job breaking down maths and complex concepts like a story!”
“I would like to thank you for your work. Specifically, the chapter: 'LIME Explainability: Complete Guide to Local Interpretable Model-Agnostic Explanations' was extremely helpful for me in understanding LIME and communicating to colleagues. Your work is inspiring and valuable. Thank you for doing it.”
“I am enjoying reading your writings. I appreciate the effort you put into it. I want to drop a note of thanks. Thank you!”
“Your writing feels like a treasure trove of great material. So, thank you for compiling everything on your website!”
“I found your book Language AI Handbook's explanations of Transformer architectures and production deployment strategies to be exceptionally clear and insightful”
“Thanks so much for establishing such an amazing community. I appreciate the opportunity to be a learner and a contributor to the community.”
“Thank you Michael. It's top!”
“I came across your collection of online handbooks covering quantitative finance, AI, and machine learning. I wanted to reach out and thank you for making such high-quality material openly available. They are incredibly helpful for my CFA Level II review and for brushing up ahead of ML interviews.”
“I came across one of your books, Language AI Handbook, and man, I'm loving it. I've read many books but have never found such a comprehensive book on large models that covers everything from start to finish with such depth.”
“I just came across your website where your books are available for free to read. Thank you so much for making this available to the public. This will help a lot of people.”
“Thanks for putting cool stuff out into the world.”
“I read a few of your blogs on NLP and oh WOW, they really are something, now I exactly know how all of the terms Entropy, Cross-entropy and how it is connected to Perplexity, will definitely read more of your blogs.”
Read all 12 reader reviews
“I would like to thank you for your work. Specifically, the chapter: 'LIME Explainability: Complete Guide to Local Interpretable Model-Agnostic Explanations' was extremely helpful for me in understanding LIME and communicating to colleagues. Your work is inspiring and valuable. Thank you for doing it.”
“Really appreciate the way you have explained the concepts. Simple to understand, yet builds up knowledge exponentially, all while reinforcing it with examples repeatedly. I've read the last two chapters `74.Self-attention & 75.Q,K,V` and I've to admit, my concepts have never been clearer. You really are doing an amazing job breaking down maths and complex concepts like a story!”
“Thanks so much for establishing such an amazing community. I appreciate the opportunity to be a learner and a contributor to the community.”
“I just came across your website where your books are available for free to read. Thank you so much for making this available to the public. This will help a lot of people.”
“I am enjoying reading your writings. I appreciate the effort you put into it. I want to drop a note of thanks. Thank you!”
“Thank you Michael. It's top!”
“Thanks for putting cool stuff out into the world.”
“I found your book Language AI Handbook's explanations of Transformer architectures and production deployment strategies to be exceptionally clear and insightful”
“Your writing feels like a treasure trove of great material. So, thank you for compiling everything on your website!”
“I came across your collection of online handbooks covering quantitative finance, AI, and machine learning. I wanted to reach out and thank you for making such high-quality material openly available. They are incredibly helpful for my CFA Level II review and for brushing up ahead of ML interviews.”
“I read a few of your blogs on NLP and oh WOW, they really are something, now I exactly know how all of the terms Entropy, Cross-entropy and how it is connected to Perplexity, will definitely read more of your blogs.”
“I came across one of your books, Language AI Handbook, and man, I'm loving it. I've read many books but have never found such a comprehensive book on large models that covers everything from start to finish with such depth.”
Exact contents
16 chapters across 2 parts
Every linked chapter is free to read online. 5 further chapters are planned; unlinked titles are still in progress.
Part XXXVII: Alignment and RLHF
- 01Alignment Problem
Covers alignment definition, helpfulness vs harmlessness, alignment challenges, alignment approaches overview.
- 02Human Preference Data
Covers preference collection UI, comparison design, annotator guidelines, preference data quality.
- 03Bradley-Terry Model
Covers pairwise comparison model, preference probability, Bradley-Terry likelihood, preference strength.
- 04Reward Modeling
Covers reward model architecture, preference loss function, reward model training, reward model evaluation.
- 05Reward Hacking
Covers reward hacking examples, distribution shift, over-optimization, reward hacking mitigation.
- 06Policy Gradient Methods
Covers policy definition, REINFORCE algorithm, policy gradient derivation, variance reduction.
- 07PPO Algorithm
Covers clipped objective, PPO derivation, trust region intuition, PPO implementation.
- 08PPO for Language Models
Covers LLM as policy, action space (tokens), reward assignment, KL penalty importance.
- 09RLHF Pipeline
Covers SFT stage, reward model training, PPO fine-tuning, RLHF hyperparameters, RLHF debugging.
- 10KL Divergence Penalty
Covers KL penalty motivation, KL coefficient selection, adaptive KL, KL effects on training.
- 11DPO Concept
Covers DPO motivation, removing reward model, DPO intuition, DPO benefits.
- 12DPO Derivation
Covers DPO from RLHF objective, optimal policy derivation, DPO loss function, DPO as classification.
- 13DPO Implementation
Covers DPO data format, DPO loss computation, DPO training procedure, DPO hyperparameters.
- 14DPO Variants
Covers IPO formulation, KTO for unpaired feedback, ORPO, cDPO, comparing alignment methods.
- 15RLAIF
Covers AI as annotator, constitutional AI principles, AI preference generation, RLAIF scalability.
- 16Iterative Alignment
Covers iterative DPO, online preference learning, self-improvement loops, alignment stability.
Part XXXVIII: Scalable Oversight and Safety Training
- 01Constitutional and Principle-Based Alignment · In progress
Written principles, critique and revision, AI feedback, evaluation, and failure modes.
- 02Scalable Oversight · In progress
Decomposition, debate, recursive supervision, weak-to-strong generalization, and oversight bottlenecks.
- 03Process Supervision and Behavioral Specifications · In progress
Supervising intermediate behavior, specifying policies, resolving conflicts, and testing adherence.
- 04Alignment Data Quality and Rater Populations · In progress
Rater disagreement, cultural pluralism, annotator effects, preference aggregation, and auditability.
- 05Alignment Robustness and Distribution Shift · In progress
Sycophancy, reward tampering, jailbreak pressure, deployment drift, and adversarial evaluation.
Choose the edition that fits your next step
Volume 12: Alignment, Preferences, and Oversight
510 pages focused on this subject, with future updates included.
$24
Choose this volumeThe complete Language AI Handbook library
All-volume access with separate volume PDFs and future updates. No combined PDF. Includes future volumes as they become available.
$199
Get all-volume accessExplore the other volumes
Volume 1 · 570 pages
Language Data and Classical NLP
Text as Data · Classical Text Representations · Distributional Semantics
Volume 2 · 790 pages
Embeddings, Tokenization, and Sequence Labeling
Word Embeddings · Subword Tokenization · Sequence Labeling
Volume 3 · Coming soon
Linguistic Structure and Meaning
Morphology · Syntax · Semantics · Discourse · Dialogue
Volume 4 · 920 pages
Neural NLP and Sequence Models
Neural Network Foundations · Recurrent Neural Networks · Sequence-to-Sequence
Volume 5 · 860 pages
Transformers from First Principles
Self-Attention · Positional Encoding · Transformer Blocks · Transformer Architectures
Volume 6 · 510 pages
Efficient, Long-Context, and Alternative Architectures
Efficient Attention · Long Context
Volume 7 · 760 pages
Pre-training Data and Scaling
Data Curation · Pre-training Objectives · Scaling Laws
Volume 8 · 440 pages
Encoder, Multilingual, and Translation Models
BERT and Variants · Encoder-Decoder Models
Volume 9 · 730 pages
Decoder Language Models and Capabilities
GPT Architecture · Modern Decoder Models · Emergent Capabilities
Volume 10 · 920 pages
Large-Scale Training and Sparse Models
Training Infrastructure · Training Optimization · Mixture of Experts
Volume 11 · 730 pages
Fine-tuning and Efficient Adaptation
Fine-tuning Fundamentals · Parameter-Efficient Fine-tuning · Instruction Tuning
Volume 13 · 220 pages
Reasoning and Verifiable Learning
Reasoning
Volume 14 · 640 pages
Efficient Inference and On-Device Models
Model Compression · Inference Optimization
Volume 15 · 600 pages
Retrieval, Knowledge, and Memory
Retrieval-Augmented Generation · Continual Learning
Volume 16 · 510 pages
Agents and AI Coding Systems
Code Generation · Tool Use and Agents
Volume 17 · 540 pages
Speech, Multimodal, and Embodied Language
Multimodal Models · Speech and Audio
Volume 18 · 950 pages
Evaluation, Factuality, and Reliability
Evaluation Fundamentals · Benchmark Evaluation · Human and Model Evaluation · Hallucination and Factuality
Volume 19 · 760 pages
Responsible, Secure, and Interpretable Language AI
Bias and Fairness · Interpretability · Safety and Security
Volume 20 · 510 pages
Applications, Production, and Compound AI Systems
LLM Applications · Production Systems
Author and edition details
About the author and this edition
Michael Brenndoerfer
Michael has spent more than a decade working across software engineering, data, AI, and business. He writes to understand difficult ideas more deeply and to share what he learns in a clear, practical way.
- Edition
- First edition
- Published
- Last reviewed
Version history
Kept current, not frozen in time
Library access includes every future volume and update. When a volume changes, its updated PDF appears in My books at no extra cost.
Current release
Edition 2026.08.0
Initial volume-library release with 19 written PDFs, a Volume 3 placeholder, and no combined edition.