Homework

Homework Guidelines

Each week has one homework, tied to that week’s topic (some weeks, it may be related to your final project instead). Each class day also has a short set of reading questions on that day’s assigned readings, due before class — see the Weekly Cadence for the schedule. Homework and readings are distributed via Google Drive; both are handed in on Canvas.

Homework is one open-ended question — not a “learn to code this” exercise. The point is to practice reasoning about an approach and cheaply testing it, not to produce polished, production-quality code.

Work through it in three steps:

1. Think it through yourself first. Before looking anything up, reason about how you would address the question. Write this down — it’s the part that’s actually being assessed.
2. See how others have addressed it. Look at the relevant prior work or literature. How does it compare to your own idea? What did you miss, or get right?
3. Code up a toy example to test it. Use Claude, GPT, or your tool of choice to quickly implement a small-scale version of your approach — a couple hours of work, not a full system. See what happens.

A few things to keep in mind:

• Using an LLM to write the code is expected and encouraged — coding assistance is fine for anything in this course. The reasoning in steps 1 and 2 is the part you should do yourself.
• The best outcome is often a surprising failure, not a clean success. If your first approach doesn’t work the way you expected, that’s usually more interesting than if it does — explain what went wrong and why. Homework is graded on the quality of your reasoning and diagnosis, not on whether your first idea worked.
• Keep the toy example small. A few hours of work on a scaled-down version of the problem is enough to get a real answer; you don’t need a large dataset or a long training run.

What to Submit

A one-page write-up, plus your code:

• Your initial approach (step 1): a paragraph or two on how you’d address the question, written before you looked anything up.
• What you found in prior work (step 2): a paragraph on how it compares to your own idea — what you got right, what you missed.
• What you tried and what happened (step 3): a short description of your toy experiment, 1-2 supporting plots/numbers if relevant, and — most importantly — your diagnosis of why it worked or didn’t.
• Code: a link to a notebook or script. It should run, but it doesn’t need to be clean — we’re reading it to check what you actually tried, not reviewing it as production code.

One page for the write-up itself (code/plots don’t count against that limit). If your diagnosis in step 3 is genuinely the interesting part, spend most of the page there rather than padding out steps 1-2.