02805 · Read this first

The new way of working

This course was rebuilt in 2026, from the ground up, around one fact: you have AI tools that can do most of what used to be homework. Here is what I changed, why I changed it, and what it asks of you.

Hear it from me, a quick personal intro. And enjoy the AI generated 80's disco goodness of the video thumbnail! Direct proof that AI can be used for terrible things as well 😅

1 · What changed

In 2025 this course graded the students on two take-home assignments and a project. Around mid-way last semester, I started taking agentic coding seriously and sat down with the best agentic tools and tried to see how well they would do on the assignments. For someone trying to build a great course for students, it was a bit disheartening. Most of the pipeline, from scraping the data through computing the statistics and creating the plots to finishing the assignments, could be more or less one-shot by a machine. Not "helped along." Solved. Gulp.

So I felt there were two things I could do. Pretend that nothing happened, maybe police harder: ban the tools, run detectors, pretend that it was still 2019.

Or try something exciting. Redesign! Accept that these tools are now part of how science gets done. AI agents are part of science, they are in the labs, they will be in your jobs. So why not rebuild the sh#t out of the course so that using agentic coding tools well would be part of the skill being taught.

Of course I chose the second option, and I have to say, I've had a blast coming up with this new thing. If you want to hear two scientists talk about how this actually changes the daily work, my friend YY Ahn and I recorded a podcast episode on doing research with Claude Code.

2 · The split

We use paper to ensure understanding.
We use agents to explore the world.

The whole redesign is that one split, applied consistently.

We need to change your brain, you need to know the material. Reading a degree distribution. Knowing what a null model is for. Predicting what happens to a network when a parameter moves. Having the intuition to identify a broken analysis. These are not typing skills, and so far no tool supplies them when you need them: in a meeting, refereeing a claim, deciding whether your own machine's output makes sense. So twice in the semester I ask you to sit down with pen and paper, no book, and show us the concepts are actually in there, in your brain somewhere. So this version of the course has replaced the old mandatory assignments with two tests (weeks 5 and 9).

For everything else, use your favorite machine friend. The weekly exercises, the go-nuts posts, the project: use every tool you have, and expect us to raise the bar accordingly. Each group's project crawls its own live Wikipedia domain. Deliverables are interactive, public, real. The goal is that the finished projects should look like things no student could have shipped in a semester five years ago. Because with your new tools, you can.

3 · The three modes

Every exercise in this course has a label, so you can see what the point is:

• 🧠 Learn — AI hands-off. This is test material: concepts, hand-calculations, interpretation. Stuff for the brain. Doing these with a machine is pointless. The test is where we check whether you did them.
• 🚀 Builder — full agentic mode. Point your AI agent at an ambitious goal and go. You are judged on the outcome and on your ability to defend every choice in it. Because when you demo, you stand next to your work and take questions, live. So you'll need to know how that community detection actually works.
• 🔬 Tool — the machine produced it, so you check it. Sometimes that's a model measuring something: label sentiment, extract entities, classify pages. Sometimes it's your agent writing the code behind a number. Either way, a result you haven't calibrated doesn't tell you anything. So in this mode the deliverable is never just the output, it is the check: ground-truth samples you labeled yourself, a hand calculation the code has to reproduce, agreement statistics, an error analysis.

4 · The skill underneath

Notice what the 🔬 mode is really teaching. The new scientific skill isn't prompting. It is knowing when to trust the machine. A machine that labels your data — or that writes the code measuring it — is a tool, like a thermometer or a survey: sometimes accurate, sometimes biased, always confident. Scientists have a hundred years of habits for exactly this situation: sample, compare against ground truth, quantify agreement, report the error. I aim to make you practice those habits until they are reflexes, and your project repo will contain an AI_METHODS.md documenting every place a model touched your results and how you checked it. When I or the TAs ask you about something, "the model said so" is not a method.

5 · The practical elements

Everything above is philosophy. Here is the practical plan:

• When and where. Wednesdays from 09:00, Auditorium 43, building 303A (auditorium 44 and classroom 46 are ours too when we need the space).
• How the day runs. There are no lectures to sit through. This site is the course. At 09:00 there is a short pep talk by moi, where I frame the week's ideas. After that you work through the week's page with your group at your own pace. Readings, explorables, exercises in your own notebooks, with the TAs and me circulating all day to help. Feel free to leave when the week's work is done (or catch up before next time).
• Groups of three. You work on the page alone, but alongside a 3-person group (smaller is OK, larger is not). Then the 'Go Nuts' exercise should be the group working together. You should try to form groups as early as possible, ideally in the first session or two, and keep them all semester (within reason). One important rule: everyone must be able to solve everything. Don't forget that the tests are individual, on paper, and they do not care who in your group understood it.

← Front page · Week 1 →