Further Reading: Chapter 1 — What AI Tools Actually Are (and Aren't)
The resources below are curated to deepen your understanding of the concepts in Chapter 1. They are organized by focus area and include notes on accessibility, cost, and the level of technical background they assume. All were accurate as of early 2025 — URLs and availability may change.
Section 1: What Large Language Models Actually Are
These resources address the fundamental question of how AI language models work — from accessible conceptual explanations to more technically detailed treatments.
1. "Attention Is All You Need" — Vaswani et al. (2017) Type: Academic paper | Cost: Free (arXiv) | Level: Technical / Advanced
The original paper introducing the Transformer architecture that underlies virtually all modern large language models. Reading this is not necessary for practical AI tool use, but it is the foundational document for anyone who wants to understand where this technology came from. The mathematics requires graduate-level familiarity with neural networks, but the abstract and introduction are readable and historically significant. Find it at arxiv.org by searching the title.
Best for: Developers, researchers, and technically oriented readers who want to trace the technology to its source.
2. "The Illustrated Transformer" — Jay Alammar (blog post) Type: Blog post with visuals | Cost: Free | Level: Technical / Accessible
Jay Alammar's visual explanation of how transformer models work is one of the most widely recommended introductions in the field. It walks through the attention mechanism with clear diagrams and step-by-step explanations. Some familiarity with basic machine learning concepts helps, but it is approachable for persistent non-specialists. Available at jalammar.github.io.
Best for: Readers who want a real technical understanding without requiring a mathematics PhD.
3. "What Is ChatGPT Doing... And Why Does It Work?" — Stephen Wolfram Type: Long-form essay | Cost: Free | Level: Accessible to Technical
Stephen Wolfram's extended essay explains how ChatGPT works from first principles, written for a general audience without assuming technical background. It covers token prediction, neural networks, training, and why the outputs are often surprisingly good. Long (20,000+ words) but readable and unusually precise for a popular treatment. Available at writings.stephenwolfram.com.
Best for: Readers who want a deep, accurate, accessible explanation and are willing to invest reading time.
4. "Large Language Models Explained With a Minimum of Math and Jargon" — Tim Lee and Sean Trott Type: Newsletter article | Cost: Free | Level: Beginner
A well-regarded accessible introduction from the Understanding AI newsletter. Covers the core concepts — token prediction, training, context windows, capabilities and limitations — without requiring any technical background. Available at understandingai.org.
Best for: Readers who want the clearest, most accessible explanation available and have no interest in technical details.
Section 2: AI Literacy and Critical Thinking About AI
These resources focus on how to think about AI tools — their societal implications, how to evaluate claims made about them, and how to develop the critical perspective needed to use them wisely.
5. "Weapons of Math Destruction" — Cathy O'Neil Type: Book | Cost: Paid (also available in most public libraries) | Level: Accessible
O'Neil's 2016 book predates the current AI tool wave but remains essential reading for anyone who wants to think critically about algorithmic systems. It covers how biases are encoded in automated systems, how confidence in mathematical-seeming outputs obscures underlying assumptions, and what happens when systems optimized for measurable metrics produce harmful real-world outcomes. The specific systems have changed; the analytical framework has not.
Best for: Readers who want to develop rigorous critical thinking about AI systems without getting lost in technical details.
6. "Atlas of AI" — Kate Crawford Type: Book | Cost: Paid (also in libraries) | Level: Accessible / Academic
A broader-scope examination of the material, labor, and power structures underlying AI technologies. Less focused on how models work technically and more focused on what they reflect about the world that produced them. Relevant to the chapter's discussion of AI tools as non-neutral systems reflecting the biases and power structures of their training data.
Best for: Readers interested in the social and political dimensions of AI, not just the technical ones.
7. "A Hacker's Guide to Language Models" — Jeremy Howard (YouTube) Type: Video / lecture | Cost: Free | Level: Technical
A 90-minute YouTube lecture by fast.ai co-founder Jeremy Howard covering how language models work, how to use them effectively, and common misconceptions. Technically informed but accessible to developers and analytically-oriented practitioners. Search "Jeremy Howard hacker's guide language models" on YouTube.
Best for: Developers and technical practitioners who want a well-structured, practitioner-focused explanation.
8. "The Alignment Problem" — Brian Christian Type: Book | Cost: Paid (also in libraries) | Level: Accessible
Christian's book covers the challenge of building AI systems that actually do what we want — and why this is harder than it sounds. Relevant to this chapter's discussion of AI tools as systems that generate plausible outputs without genuine understanding of what we need. More conceptual than technical, extensively researched, and well-written.
Best for: Readers who want a thorough treatment of why AI systems behave unexpectedly and what that means for how we use them.
Section 3: Hallucination, Accuracy, and Verification
These resources address the hallucination problem specifically — what it is, how it has been studied, and what can be done about it.
9. "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" — Bender et al. (2021) Type: Academic paper | Cost: Free (ACL Anthology) | Level: Academic / Accessible
An influential paper examining the risks of large language models, including the generation of fluent but misleading text, the reinforcement of existing biases, and the environmental costs of training large models. The "stochastic parrot" framing — the idea that language models can produce human-seeming language without understanding — remains a useful conceptual tool. Available at aclanthology.org.
Best for: Readers who want to understand the scholarly critique of language model capabilities and the "understanding vs. fluency" debate.
10. "TruthfulQA: Measuring How Models Mimic Human Falsehoods" — Lin et al. (2022) Type: Academic paper | Cost: Free (arXiv) | Level: Technical / Research
A research paper studying how language models reproduce false beliefs that are common among humans — the kinds of plausible-sounding wrong answers that appear frequently in training data. Shows that larger models can actually become more confident about wrong answers that appear frequently in human-generated text. Empirically grounded and readable for persistent non-specialists.
Best for: Readers who want empirical evidence for the hallucination and confidence-miscalibration problems discussed in Chapter 1.
11. "Evaluating Language Models Is Harder Than It Looks" — Various authors (multiple resources) Type: Multiple articles and posts | Cost: Free | Level: Accessible to Technical
Search for this topic on the Hugging Face blog, Towards Data Science, and academic AI forums. The challenge of evaluating AI model accuracy is itself a rich topic — standard benchmarks can be gamed, test sets get contaminated by training data, and performance on one task predicts little about performance on another. Understanding this helps calibrate skepticism about AI capability claims.
Best for: Readers who want to understand why AI capability claims should be treated carefully and how to evaluate them.
Section 4: Practical AI Literacy for Professionals
These resources focus on the practical dimension — how to work with AI tools effectively in professional contexts.
12. "Co-Intelligence: Living and Working with AI" — Ethan Mollick Type: Book | Cost: Paid | Level: Accessible / Practitioner
Mollick, a Wharton professor who has studied AI in educational and professional contexts extensively, offers a practical and intellectually honest guide to working with AI tools. He is neither a techno-utopian nor a doom-sayer, and his grounded, research-informed perspective is valuable. Directly relevant to the practical themes of this chapter and book.
Best for: Professionals who want a thoughtful, research-backed practical guide to AI tool use — especially in knowledge work.
13. "The Pragmatic Programmer's Guide to AI Tools" — Various Authors (ongoing blog posts and newsletters) Type: Ongoing publications | Cost: Free and Paid options | Level: Technical / Practitioner
Several high-quality blogs and newsletters cover practical AI tool use for developers specifically. Notable options include Simon Willison's blog (simonwillison.net) for technically precise, experience-based writing on language models; the Hacker News AI thread archives for practitioner community perspective; and the AI Snake Oil newsletter (aisnakeoil.com) for rigorous debunking of overblown AI claims.
Best for: Developers and technically-minded practitioners who want ongoing, practical, critically-minded coverage.
14. "Sparks of Artificial General Intelligence: Early Experiments with GPT-4" — Bubeck et al. (2023) Type: Research paper | Cost: Free (arXiv) | Level: Technical / Research
A Microsoft Research paper studying GPT-4's capabilities across a wide range of tasks. Notable for being both genuinely impressed by what the model can do and rigorous about the limits. The tension between "remarkably capable" and "not actually reasoning" is examined with more nuance than most popular coverage. Long, but readable sections for non-specialists.
Best for: Readers who want a serious, empirically grounded treatment of what current language models can and cannot do, from researchers who studied them intensively.
15. "AI Literacy: What It Is and How to Develop It" — Various organizations (ongoing) Type: Courses and frameworks | Cost: Free options widely available | Level: Beginner to Intermediate
AI literacy has become a recognized educational concept with multiple curricula and frameworks developed by organizations including MIT Media Lab, Google's "AI Literacy" resources, and various university programs. Search "AI literacy course" for current offerings. Most introductory-level courses are free and cover foundational concepts without requiring technical background.
Best for: Readers who want structured, progressive AI literacy development beyond what a single book provides.
A Note on Staying Current
The field of AI develops rapidly, and any resource published more than two years ago should be read with an awareness that the specific tools, capabilities, and benchmarks described may have changed substantially. What does not change quickly is the foundational architecture and the fundamental limitations discussed in this chapter. Token prediction, training cutoffs, the absence of genuine understanding, the hallucination problem — these are properties of the current generation of tools, not bugs that will be patched next month.
For staying current with AI tool developments as a practitioner, high-quality options include:
- Following researchers directly on academic preprint servers (arXiv.org)
- Newsletter subscriptions from practitioners you trust (evaluate sources carefully — the field attracts hype)
- Hands-on experimentation with new tools as they release, evaluated against your own specific use cases rather than demo scenarios
The habit of applied skepticism — testing claims against your own experience rather than accepting them from marketing or media coverage — is the most durable AI literacy skill of all.