How Large Language Models Work

Learn how large language models like GPT and Gemini work under the hood in plain English.

How Large Language Models Work translates years of expert research on Large Language Models into a readable, focused introduction to working with these amazing systems. It explains clearly how LLMs function, introduces the optimization techniques to fine-tune them, and shows how to create pipelines and processes to ensure your AI applications are efficient and error-free.

In How Large Language Models Work you will learn how to:

• Test and evaluate LLMs
• Use human feedback, supervised fine-tuning, and Retrieval Augmented Generation (RAG)
• Reducing the risk of bad outputs, high-stakes errors, and automation bias
• Human-computer interaction systems
• Combine LLMs with traditional ML

How Large Language Models Work is authored by top machine learning researchers at Booz Allen Hamilton, including researcher Stella Biderman, Director of AI/ML Research Drew Farris, and Director of Emerging AI Edward Raff. They lay out how LLM and GPT technology works in plain language that’s accessible and engaging for all.

About the Technology

Large Language Models put the “I” in “AI.” By connecting words, concepts, and patterns from billions of documents, LLMs are able to generate the human-like responses we’ve come to expect from tools like ChatGPT, Claude, and Deep-Seek. In this informative and entertaining book, the world’s best machine learning researchers from Booz Allen Hamilton explore foundational concepts of LLMs, their opportunities and limitations, and the best practices for incorporating AI into your organizations and applications.

About the Book

How Large Language Models Work takes you inside an LLM, showing step-by-step how a natural language prompt becomes a clear, readable text completion. Written in plain language, you’ll learn how LLMs are created, why they make errors, and how you can design reliable AI solutions. Along the way, you’ll learn how LLMs “think,” how to design LLM-powered applications like agents and Q&A systems, and how to navigate the ethical, legal, and security issues.

What’s Inside

• Customize LLMs for specific applications
• Reduce the risk of bad outputs and bias
• Dispel myths about LLMs
• Go beyond language processing

About the Readers

No knowledge of ML or AI systems is required.

About the Author

Edward Raff, Drew Farris and Stella Biderman are the Director of Emerging AI, Director of AI/ML Research, and machine learning researcher at Booz Allen Hamilton.

Table of Contents

1 Big picture: What are LLMs?
2 Tokenizers: How large language models see the world
3 Transformers: How inputs become outputs
4 How LLMs learn
5 How do we constrain the behavior of LLMs?
6 Beyond natural language processing
7 Misconceptions, limits, and eminent abilities of LLMs
8 Designing solutions with large language models
9 Ethics of building and using LLMs

About the Author

Edward Raff is a Director of Emerging AI at Booz Allen Hamilton, where he leads the machine learning research team. He has worked in healthcare, natural language processing, computer vision, and cyber security, among fundamental AI/ML research. The author of Inside Deep Learning, Dr. Raff has over 100 published research articles at the top artificial intelligence conferences. He is the author of the Java Statistical Analysis Tool library, a Senior Member of the Association for the Advancement of Artificial Intelligence, and twice chaired the Conference on Applied Machine Learning and Information Technology and the AI for Cyber Security workshop. Dr. Raff's work has been deployed and used by anti-virus companies all over the world.

Drew Farris is a professional software developer and technology consultant whose interests focus on large scale analytics, distributed computing and machine learning. Previously, he worked at TextWise where he implemented a wide variety of text exploration, management and retrieval applications combining natural language processing, classification and visualization techniques. He has contributed to a number of open source projects including Apache Mahout, Lucene and Solr, and holds a master's degree in Information Resource Management from Syracuse University's iSchool and a B.F.A in Computer Graphics.

Stella Biderman is a machine learning researcher at Booz Allen Hamilton and the executive director of the non-profit research center EleutherAI. She is a leading advocate for open source artificial intelligence and has trained many of the world's most powerful open source artificial intelligence algorithms. She has a master's degree in computer science from the Georgia Institute of Technology and degrees in Mathematics and Philosophy from the University of Chicago.

Ebook License

End-User Warranty And License Agreement

1. Grant Of License

Manning Has Authorized The Download By You Of An Unrestricted Number Of Copies Of The Electronic Book (Ebook) In Any Of The Available Formats. Manning Grants You A Nonexclusive, Nontransferable License To Use The Ebook According To The Terms And Conditions Herein. This License Agreement Permits You To Install The Ebook On Any And All Your Devices For Your Personal Use Only.

2. Restrictions

You Shall Not: (1) Share, Resell, Rent, Assign, Timeshare, Distribute, Or Transfer All Or Part Of The Ebook Or Any Rights Granted Hereunder To Any Other Person; (2) Duplicate The Ebook, Except For A Single Backup Or Archival Copy; (3) Remove Any Proprietary Notices, Labels, Or Marks From The Ebook; (4) Transfer Or Sublicense Title To The Ebook To Any Other Party.

3. Intellectual Property Protection

The Ebook Is Owned By Manning And Is Protected By United States And International Copyright And Other Intellectual Property Laws. Manning Reserves All Rights In The Ebook Not Expressly Granted Herein. This License And Your Right To Use The Ebook Terminate Automatically If You Violate Any Part Of This Agreement. In The Event Of Termination, You Must Remove The Original And Any Copies Of The Ebook From All Your Devices.

4. Source Code Supplementary Material

Any Source Code Files Provided As A Supplement To The Book Are Freely Available To The Public For Download. Reuse Of The Code Is Permitted, In Whole Or In Part, Including The Creation Of Derivative Works, Provided That You Acknowledge That You Are Using It And Identify The Source: Title, Publisher And Year.

5. Limited Warranty

Manning Warrants That The Ebook Files, A Copy Of Which You Are Authorized To Download, Are Free From Defects In The Operational Sense That They Can Be Read By A Pdf Reader Or Epub Reader, Or Other. Except For This Express Limited Warranty, Manning Makes And You Receive No Warranties, Express, Implied, Statutory Or In Any Communication With You, And Manning Specifically Disclaims Any Other Warranty Including The Implied Warranty Of Merchantability Or Fitness Or A Particular Purpose. Manning Does Not Warrant That The Operation Of The Ebook Will Be Uninterrupted Or Error Free. If The Ebook Was Purchased In The United States, The Above Exclusions May Not Apply To You As Some States Do Not Allow The Exclusion Of Implied Warranties. In Addition To The Above Warranty Rights, You May Also Have Other Rights That Vary From State To State.

6. Limitation Of Liability

In No Event Will Manning Be Liable For Any Damages, Whether Arising For Tort Or Contract, Including Loss Of Data, Lost Profits, Or Other Special, Incidental, Consequential, Or Indirect Damages Arising Out Of The Use Or Inability To Use The Ebook.

7. General

This Agreement Constitutes The Entire Agreement Between You And Manning And Supersedes Any Prior Agreement Concerning The Ebook. This Agreement Is Governed By The Laws Of The State Of New York.