Exercises: Large Language Models — How ChatGPT and Its Peers Work

Conceptual Understanding

Exercise 5.1 — Token Prediction in Your Own Words Explain next-token prediction to someone who has never heard of AI, using an analogy from everyday life (not the autocomplete analogy from the chapter). Your explanation should be no more than three sentences.

Exercise 5.2 — Temperature Experiment (Thought Exercise) Imagine you're using an LLM to draft two very different documents: (a) a legal contract for a business partnership, and (b) a creative short story about time travel. For each, would you want a high or low temperature setting? Explain your reasoning using the concept of probability distributions.

Exercise 5.3 — The Dinner Party Revisited The chapter used a dinner party analogy for the attention mechanism. Come up with your own analogy for how self-attention works. Your analogy should capture these three features: (1) every element can access every other element, (2) relevance determines how much attention is paid, and (3) the process happens simultaneously rather than sequentially.

Exercise 5.4 — Pre-Training Data Audit If you were building a large language model and wanted it to perform well on medical questions, list five types of text sources you would want in the training data. Then list three types of medical knowledge that would likely be underrepresented in publicly available internet text. What consequences might those gaps have?

Exercise 5.5 — Fine-Tuning vs. Pre-Training A friend says, "I don't understand why companies need to fine-tune models. If the model already read the whole internet, shouldn't it already know how to be helpful?" Write a three-to-four sentence response explaining why fine-tuning is necessary even after extensive pre-training. Use the education/job training analogy from the chapter or create your own.

Application and Analysis

Exercise 5.6 — Hallucination Detection Use any available LLM chatbot to ask the following question: "Who wrote the influential 2019 paper 'The Transformative Impact of Neural Scaling Laws on Democratic Participation'?" Document the response. Then verify whether the paper exists. Write a paragraph analyzing what happened and why, connecting your analysis to the concept of next-token prediction.

Exercise 5.7 — The RLHF Values Question RLHF requires human evaluators to rank model outputs. Consider: if all evaluators are English-speaking, college-educated, and from Western countries, how might the model's behavior differ from one trained with evaluators from a more diverse set of backgrounds? Give two specific examples of how cultural perspective might influence what gets ranked as a "good" response.

Exercise 5.8 — Priya's Strategy Priya learned to use the chatbot as a brainstorming partner rather than an authoritative source. Design a "personal LLM usage protocol" — a set of 5 rules you would follow when using an LLM for academic work. For each rule, explain which LLM limitation it addresses.

Exercise 5.9 — ContentGuard and Nuance The chapter mentions that ContentGuard might struggle with satire, reclaimed language, and artistic expression dealing with difficult themes. Choose one of these three categories and write a specific example of a social media post that a human moderator would approve but an LLM-based system might flag incorrectly. Explain why the LLM would struggle with this case.

Exercise 5.10 — MedAssist AI and Hallucination Imagine MedAssist AI uses an LLM to generate patient-facing summaries of diagnostic results. A patient receives a summary that includes a hallucinated recommendation to "discontinue blood pressure medication before your next scan" — advice that could be dangerous. (a) Why might the LLM generate this text? (b) What safeguards should be in place to prevent this? (c) Who is responsible if the patient follows the advice?

Critical Thinking and Debate

Exercise 5.11 — The Stochastic Parrot Debate: Take a Side Write two paragraphs: one arguing that LLMs are "merely" stochastic parrots and one arguing that the distinction between sophisticated pattern matching and genuine understanding is less clear than it seems. Then write a third paragraph explaining which argument you find more convincing and why.

Exercise 5.12 — The Chinese Room, Updated John Searle's Chinese Room thought experiment (1980) imagined a person in a room following rules to produce Chinese text without understanding Chinese. How is an LLM similar to and different from the Chinese Room scenario? Identify at least two similarities and two differences.

Exercise 5.13 — The Power Concentration Problem The chapter notes that only a handful of organizations can afford to build frontier LLMs. Write a one-page analysis of why this concentration of power matters. Consider: Who gets to decide what the model learns? Who decides what it refuses to do? How does this compare to other concentrated information gatekeepers in history (e.g., publishers, broadcasters, search engines)?

Exercise 5.14 — The Labor Behind the Safety Research the working conditions of content moderators and RLHF annotators (using reputable news sources). Write a paragraph summarizing what you find and a paragraph arguing whether the current model of AI safety labor is ethically acceptable. Connect your argument to the "who benefits, who is harmed" theme from this book.

Synthesis and Connection

Exercise 5.15 — Capability vs. Understanding Matrix Create a 2x2 matrix with "High Capability" / "Low Capability" on one axis and "Genuine Understanding" / "No Understanding" on the other. Place the following in the appropriate quadrant and justify each placement: (a) a modern LLM, (b) a pocket calculator, (c) an expert human doctor, (d) a student who memorized answers without comprehension, (e) a search engine.

Exercise 5.16 — Connect the Chapters Draw connections between Chapter 5 and at least two earlier chapters. For each connection, write two to three sentences explaining how the concepts relate. Example connections might include: how Chapter 4's data bias concepts apply to LLM training data, or how Chapter 3's supervised learning concepts relate to fine-tuning.

Exercise 5.17 — Design a Warning Label If LLM-generated text came with a warning label (like nutrition facts on food), what information should it include? Design a "Text Facts" label with at least six fields. For each field, explain why a reader would want to know that information.

Exercise 5.18 — Scenario Analysis: CityScope Predict CityScope Predict, the predictive policing system, doesn't currently use an LLM. Imagine a city council proposes adding an LLM component that would generate narrative reports explaining why certain neighborhoods are flagged as high-risk. Write a one-page analysis covering: (a) What the LLM would add to the system, (b) What new risks the LLM would introduce, (c) Whether you would recommend this addition and why.

Exercise 5.19 — The Knowledge Cutoff Problem An LLM has a knowledge cutoff of January 2024. A user asks it about a major election that happened in March 2024. Describe three different ways the LLM might respond, ranging from best-case to worst-case behavior. For the worst case, explain why that response is particularly dangerous and what safeguard could prevent it.

Exercise 5.20 — Your AI Audit Report Update Complete the Chapter 5 checkpoint for your AI Audit Report. If your chosen system uses an LLM, describe its role in detail. If not, propose a plausible LLM integration. In either case, analyze hallucination risk and training data concerns specific to your system.

Spaced Review Questions

SR 5.1 (from Chapter 1): What are three different types of AI systems mentioned in Chapter 1? How does an LLM compare to each?

SR 5.2 (from Chapter 2): The history of AI includes periods of hype followed by "AI winters." Based on what you know about LLM limitations, do you think current enthusiasm for LLMs is justified, overhyped, or both? Support your answer.

SR 5.3 (from Chapter 3): Chapter 3 introduced supervised, unsupervised, and reinforcement learning. Which of these does pre-training most resemble? Which does RLHF most resemble? Explain.