Syllabus
We follow an applied-ML project pipeline (reading, building, data and evaluation, and writing up). Each module asks how to do that part well with AI.
For each AI idea or practice we introduce, we also ask four questions: What problem was it solving? What is the proposed solution? Which older ideas help explain it? What is genuinely new?
The first three answers teach a craft to run or a system to reason about, or name why the model, the data, or the evaluation breaks. The fourth is about thinking in deltas, which outlasts the tools that prompt it.
Alongside the lectures run the two deliverables: a semester-long project built around one small research question, and a personal portfolio of a SKILL.md bundle and a weekly agency log.
Schedule
Lectures Tuesday, labs Thursday, starting September 15.
| Topic | Lec | Lab |
|---|---|---|
| Module 1: framing & reading | ||
| what this course is | Sep 15 | Sep 17 |
| reading matters | Sep 22 | Sep 24 |
| Module 2: build & the harness | ||
| why still build | Sep 29 | Oct 1 |
| specs before prompts | Oct 6 | Oct 8 |
| git makes a diff | Oct 20 | Oct 22 |
| the abstraction ladder (of programming) | Oct 27 | Oct 29 |
| Module 3: data & evaluation | ||
| outsourced data | Nov 3 | Nov 5 |
| the machinery beyond vanilla transformers | Nov 10 | Nov 12 |
| evals as experimental design | Nov 17 | Nov 19 |
| auditing the artifact and the judge | Nov 24 | Dec 3 |
| Module 4: reflection & writing | ||
| what the (human) tutor had | Dec 1 | Dec 10 |
| closing arguments | Dec 8 | Dec 10 |
- No lecture or lab the week of October 13, due to the Monday schedule shift.
- The last three labs run later than their lectures, due to the Thanksgiving holiday on November 26.
- The final two labs share the December 10 session.
Learning objectives
The takeaways stack in order:
- how to use AI tools;
- where and why these tools work well;
- where and why these tools fail by construction;
- how to orchestrate work around the tools based on that why-and-why-not understanding.
More concretely, students should be able to:
- specify and decompose work for an AI agent: e.g., specs first, scoped increments, skills as reusable contracts;
- keep the work inspectable: e.g., version history and provenance for AI-co-authored code, context held to a budget;
- form and test mechanism hypotheses for model failures: e.g., training objective, context limits, sampling;
- judge data and design evaluations: e.g., contamination checks, sample size, auditing the judge itself;
- check an AI-assisted reading against the source and build a map of the field: e.g., what the paper’s delta is and relative to what;
- write up a result as claim, evidence, and limits, and edit AI-drafted prose without losing authorship;
- justify their delegate, verify, redirect, and decline calls with evidence: e.g., a transcript moment, a commit.
Page updated August 31, 2026.