Image by Author

# Introduction

I know a lot of people want to study LLMs deeply, and although courses and articles are great for getting broader knowledge, one really needs to refer to books for an in-depth understanding. Another thing I personally like about books is their structure. They have an order that is more intuitive and cohesive compared to courses which can sometimes feel all over the place. With this motivation, we’re starting a new series for our readers to recommend 5 FREE but totally worth it books for different roles. So, if you’re serious about understanding how large language models (LLMs) really work, then here are my recommendations for 5 FREE books that you should start with.

# 1. Foundations of Large Language Models

Published in early 2025, Foundations of Large Language Models is one of the most well-structured and conceptually clear books written for anyone who wants to actually understand how LLMs are built, trained, and aligned. The authors (Tong Xiao & Jingbo Zhu) are both well-known figures in natural language processing (NLP). Instead of rushing through every new architecture or trend, they carefully explain the core mechanisms behind modern models like GPT, BERT, and LLaMA.

The book emphasizes foundational thinking: what pre-training actually means, how generative models function internally, why prompting strategies matter, and what “alignment” really involves when humans try to fine-tune machine behavior. I think it’s a thoughtful balance between theory and implementation, designed both for students and practitioners who want to build strong conceptual grounding before starting experimentation.

// Overview of Outline

1. Pre-training (overview, different paradigms, bert, practical aspects of adapting and applying pre-trained models, etc.)
2. Generative models (decoder-only transformers, data preparation, distributed training, scaling laws, memory optimization, efficiency strategies, etc.)
3. Prompting (principles of good prompt design, advanced prompting methods, techniques for optimizing prompts)
4. Alignment (LLM alignment and RLHF, instruction tuning, reward modeling, preference optimization)
5. Inference (guidance on decoding algorithms, evaluation metrics, efficient inference methods)

# 2. Speech and Language Processing

If you want to understand NLP and LLMs deeply, [Speech and Language Processing](http://<a href=) by Daniel Jurafsky and James H. Martin is one of the best resources. The 3rd edition draft (August 24, 2025 release) is fully updated to cover modern NLP, including Transformers, LLMs, automatic speech recognition (Whisper), and text-to-speech systems (EnCodec & VALL-E). Jurafsky and Martin are leaders in computational linguistics, and their book is widely used in top universities.

It provides a clear, structured approach from the basics like tokens and embeddings to advanced topics such as LLM training, alignment, and conversation structure. The draft PDF is freely available, making it both practical and accessible.

// Overview of Outline

• Volume I: Large Language Models

• Chapters 1–2: Introduction, words, tokens, and Unicode handling

• Chapters 3–5: N-gram LMs, Logistic Regression for text classification, and vector embeddings

• Chapters 6–8: Neural networks, LLMs, and Transformers — including sampling and training techniques

• Chapters 9–12: Post-training tuning, masked language models, IR & RAG, and machine translation

• Chapter 13: RNNs and LSTMs (optional ordering for learning sequence models)

• Chapters 14–16: Phonetics, speech feature extraction, automatic speech recognition (Whisper), and text-to-speech (EnCodec & VALL-E)

• Volume II: Annotating Linguistic Structure

• Chapters 17–25: Sequence labeling, POS & NER, CFGs, dependency parsing, information extraction, semantic role labeling, lexicons, coreference resolution, discourse coherence, and conversation structure

# 3. How to Scale Your Model: A Systems View of LLMs on TPUs

Training LLMs can be difficult because the numbers are huge, the hardware is complex, and it’s hard to know where the bottlenecks are. How to Scale Your Model: A Systems View of LLMs on TPUs takes a very practical, systems-oriented approach to explain the performance side of LLMs like how Tensor Processing Units (TPUs) (and GPUs) work under the hood, how these devices communicate, and how LLMs actually run on real hardware. It also covers parallelism strategies for both training and inference to efficiently scale models at massive sizes.

This resource stands out because the authors have actually worked on production-grade LLM systems themselves at Google, so they share their learnings.

// Overview of Outline

• Part 0: Rooflines (understanding hardware constraints: flops, memory bandwidth, memory)
• Part 1: TPUs (how TPUs work and network together for multi-chip training)
• Part 2: Sharding (matrix multiplication, TPU communication costs)
• Part 3: Transformer math (calculating flops, bytes, and other critical metrics)
• Part 4: Training (parallelism strategies: data parallelism, fully-sharded data parallelism (FSDP), tensor parallelism, pipeline parallelism)
• Part 5: Training LLaMA (practical examples of training llama 3 on TPU v5p; cost, sharding, and size considerations)
• Part 6: Inference (latency considerations, efficient sampling and accelerator utilization)
• Part 7: Serving LLaMA (serving llama 3-70b models on TPU v5e; kv caches, batch sizes, sharding, and production latency estimates)
• Part 8: Profiling (practical optimization using XLA compiler and profiling tools)
• Part 9: JAX (programming TPUs efficiently with JAX)

# 4. Understanding Large Language Models: Towards Rigorous and Targeted Interpretability Using Probing Classifiers and Self-Rationalisation

Understanding Large Language Models: Towards Rigorous and Targeted Interpretability Using Probing Classifiers and Self-Rationalisation is not a typical textbook. It’s a doctoral thesis of Jenny Kunz from Linköping University, but it covers such a unique aspect of LLMs that it deserves a place in this list. She explores how large language models work and how we can better understand them.

LLMs perform very well on many tasks, but it is not clear how they make their predictions. This thesis studies two ways to understand these models: looking at the internal layers using probing classifiers and examining the explanations models generate for their predictions. She also examines models that generate free-text explanations with their predictions, exploring which properties of these explanations actually help downstream tasks and which align with human intuition. This work is useful for researchers and engineers interested in creating more transparent and accountable AI systems.

// Overview of Outline

1. Understanding LLM layers with probing classifiers (analyzing information stored in each layer of the model, checking limitations of existing probing methods, creating stricter probing tests using changes in data, developing new ways to measure differences in what layers know)
2. Explaining predictions with self-rationalising models (generating text explanations along with model predictions, comparing explanations with human ratings and task performance, studying which properties make explanations useful for tasks versus easy to understand, annotating explanations for human-like features and their effects on different users)

# 5. Large Language Models in Cybersecurity: Threats, Exposure and Mitigation

LLMs are very powerful, but they can also create risks such as leaking private information, helping with phishing attacks, or introducing code vulnerabilities. Large Language Models in Cybersecurity: Threats, Exposure and Mitigation explains these risks and shows ways to reduce them. It covers real examples, including social engineering, monitoring LLM adoption, and setting up safe LLM systems.

This resource is unique because it focuses on LLMs in cybersecurity, a topic most LLM books do not cover. It is very useful for anyone who wants to understand both the dangers and protections related to LLMs.

// Overview of Outline

• Part I: Introduction (how LLMs work and how they are used, limits of LLMs and evaluation of their tasks)
• Part II: LLMs in cybersecurity (risks of private information leakage, phishing and social engineering attacks, vulnerabilities from code suggestions, LLM-assisted influence operations and web indexing)
• Part III: Tracking and forecasting exposure (trends in LLM adoption and risks, investment and insurance aspects, copyright and legal issues, monitoring new research in LLMs)
• Part IV: Mitigation (security education and awareness, privacy-preserving training methods, defenses against attacks and adversarial use, LLM detectors, red teaming, and safety standards)
• Part V: Conclusion (the dual role of LLMs in causing threats and providing defenses, recommendations for safe use of LLMs)

# Wrapping Up

All five of these books approach LLMs from very different angles: theory, linguistics, systems, interpretability, and security. Collectively, they form a complete learning path for anyone serious about learning large language models. If you liked this article, let me know in the comments section below which topics you’d like to explore further.

Kanwal Mehreen is a machine learning engineer and a technical writer with a profound passion for data science and the intersection of AI with medicine. She co-authored the ebook “Maximizing Productivity with ChatGPT”. As a Google Generation Scholar 2022 for APAC, she champions diversity and academic excellence. She’s also recognized as a Teradata Diversity in Tech Scholar, Mitacs Globalink Research Scholar, and Harvard WeCode Scholar. Kanwal is an ardent advocate for change, having founded FEMCodes to empower women in STEM fields.