Homework for Class 4
- We’re extending the Grok Bot challenge by one week. Set up at least one Grok Bot to help you out, and share it in the #homework channel by Class 4. We’ll pick the coolest Grok Bot in the cohort, and the winner gets a free month of Grok Bot.
- Make sure your homework automation posts to the #homework channel every day, and sends us an email with an update on your progress. If your automation isn’t working, ask Codex to look at past automation runs, and help you debug why the automation isn’t working.
- Run the agent design game on your goals again. Pick a problem in your life, work through it with your agent, and spend several hours iterating on the solution.
- Prepare to show your work in demos next class. Demos will be done in groups of 3-4 students, with volunteers showing the whole class something they created in the past week.
- Memorize the terms in the attached glossary sheet. We’ll do a short quiz at the start of class 4 on these terms.
Study Guide for Class 4’s Quiz Ten terms on choosing a model, long chats, and how models fail. The sheet for the Class 4 quiz. Two pages, PDF. Open ↗ (opens in a new tab)
- Send Codex the following questions, and ask it to explain them to you (and ask follow up questions) until you feel satisfied with your understanding:
- What is a model family? Which models can I choose from in Codex, and how do I decide which one to use for a task?
- What is effort? What does the model actually do when it “thinks” before answering, and when is it worth turning effort up?
- What does it mean that a model is stateless? If it remembers nothing between messages, how does it keep track of a long conversation?
- What are the smart zone and the dumb zone? Why do models get worse as a chat gets longer, and how can I tell when I’m in the dumb zone?
- What is compaction? What gets kept and what gets lost when a chat is compacted, and how do I make sure the important details survive?
- What is a subagent? When would an agent start one, and how does that keep the main chat sharp?
- What is non-determinism? Why does the same prompt give me different answers, and how can I get more consistent results?
- What is a hallucination? Why do models make things up so confidently, and how can I catch it when it happens?
- What is sycophancy? Why do models tell me what I want to hear, and how can I ask questions that get honest feedback?
- What is prompt injection? How could an email or a web page trick my agent, and what are AI companies doing to stop it?