None of the readings on this page are required, and none of them will be assessed. They’re here for the times when something in lecture or tutorial catches your interest and you want to go further than the course itself goes.
Readings are grouped by the week they connect to, so you can follow along or dip in whenever a topic grabs you.
Feel free to suggest additions on Ed β if you find something good, we’d like to share it.
Recommended book
When I’m asked, “If I read just one book about AI, what should it be?” this is the answer I give. Author Ethan Mollick teaches at Wharton Business School, and takes a very user and use-case perspective on AI that starts at a beginner-friendly level and goes surprisingly deep for a non-computer scientist. It explores practical use cases, educational impacts, and a variety of possibilities for what the future of AI could bring, with the requisite humility that nobody can know for certain.
Week 1
Welcome, Introduction to AI
- What happens when you talk to AI? β Anthropic (video). A friendly, non-technical explanation of what goes on behind a chat window.
- Transformers, explained: Understand the model behind GPT, BERT, and T5 β Google Cloud (video). Another short, clear take on transformers.
- Visualizing transformers and attention β Grant Sanderson, 3Blue1Brown (video). Long and detailed, but the animation is unusually good.
- Transformers β Dan Jurafsky and James Martin (textbook chapter). Far more depth than this course needs; look for the example of the (tired) chicken crossing the (wide) road.
- Attention Is All You Need β Ashish Vaswani et al., NeurIPS 2017 (paper, PDF). The paper that introduced the transformer. Dense, but worth a skim once the videos above have given you the shape of it.
- How random contingencies created the current AI moment β Ethan Mollick (LinkedIn post, video). A short, Connections-style history of how unrelated ideas β Boole’s logic, gaming GPUs, ImageNet, Reddit upvotes β added up to today’s AI. The video was itself made with AI, so watch it with this course’s habit of checking claims.
Week 2
Core Prompting Skills and Model/Mode Selection
- Claude System Prompts
- Anthropic publishes each model’s system prompt here in the name of transparency. They can make for interesting reading!
- Lost in the Middle: How Language Models Use Long Contexts
- Liu et. al., research paper (including Stanford reserachers like Prof. Percy Liang) documenting behavior similar to human “primary bias” and “recency bias,” and that information in the middle of the context window is devalued.
- I Love Generative AI and Hate the Companies Building It β Christina Wodtke (blog post). A look at which models are developed most ethically β another lens for choosing a model.
Week 3
Managing Context and Retrieval
- Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts β Jian Xie, Kai Zhang, Jiangjie Chen, Renze Lou, and Yu Su, ICLR 2024 (paper, PDF). What a model does when the context you give it contradicts what it learned in training: convincing evidence usually wins, but models also show a strong confirmation bias.
Week 4
Research Innovation with AI
- Using an LLM for Humanities research β Rachel Helps (blog post). A practitioner’s account of putting an LLM to work on humanities research.
- AI Legal Review Says Millions Live Under Discriminatory Local Laws β Andrew Myers, Stanford HAI (article). AI used to review legal texts at a scale no human team could manage.
Week 5
Controlling and Giving Permission with Context
- Stanford-approved connectors grid
- A grid showing which connectors are approved for which AI tools.
- Elephants, Goldfish and the New Golden Age of Software Engineering β Dave Rensin, Google (blog post). Proposes the “Elephant-Goldfish” workflow: write the design doc in one long AI conversation (the elephant), then hand it to a fresh session with no memory (the goldfish) to check that it stands on its own before any code gets written.
- Specification Gaming: How AI Can Turn Your Wishes Against You β Rational Animations (video). Short and engaging: what happens when a model does what you said rather than what you meant.
- How to explain agentic AI to your leadership team β Ashley Nicholson (LinkedIn post).
Week 6
Managing Agentic AI, Project Context, and Complex Workflows
- Practices for Governing Agentic AI Systems
- This white paper is a collaboration of Stanford HAI and Open AI, defining key risks and recommendations for agentic AI.
- NIST AI Risk Management Framework (AI RMF)
- This is the US Government’s official set of standards guiding what it means to do safe, secure, and trustworthy AI development. A document (or really, a whole set of documents, standards, playbooks for implementation, and more) that every AI leader should be familiar with.
- Responsible AI at Stanford
- Stanford’s in-house set of policies, regulations, and guidance on responsible AI use, especially as it relates to sensitive data about students, Stanford hospital patients, alumni, etc.
- Exclusive: US military had close call after using AI for false intelligence report, sources say β Katie Bo Lillis and Zachary Cohen, CNN (article). A high-stakes reminder of why AI output has to be verified.
Week 7
Eval Design
- Teaching the New Loop β Andy Hall (blog post). A Stanford instructor on teaching students to write their own evaluation rubrics for AI models and build them into practical tools.
- Universities should arm students with AI ’eval’ powers β Andy Hall, Financial Times (article; may be paywalled). The case for teaching evaluation as a core skill.
- Define success criteria and build evaluations β Anthropic (documentation). A practical guide to writing success criteria and tests for model output.
Week 8
Token and Budget Management
- Using Claude Code: Spending your effort β Thariq Shihipar (blog post). What “effort” settings actually buy you, with experiments on when low versus high effort pays off.
Week 9
Learning and Thinking with AI
- What happens to learning when AI does the thinking for us?
- Stanford Report summary of learning research insights about AI use.
- Rethinks: Defining Where We Want AI to Take Us β Matt Abrahams with Russ Altman, Think Fast Talk Smart episode 324 (podcast). A Stanford professor’s case for AI as an amplifier rather than a replacement: if you’re clear about why you’re communicating and what you’re trying to say, the tools should sharpen your message, not supplant it.
- OlmoEarth v1.3: open embeddings for every 10 m of Earth β Ai2 (Reddit post). Ai2’s open geospatial model, released with ready-made embeddings for the whole planet. Worth a look if your final project could use satellite or environmental data.