AI for Economic Research: Dynamic Models, Language, and Agents was taught by Zhigang Feng at Zhejiang University (浙江大学) as a 24-hour graduate intensive over six meetings in July 2026, from ten lectures. Announcements and Q&A ran through a course WeChat group (群聊：湖畔宏观); that group invite expired on 12 July 2026 and is not reproduced here.
| Date | Time | Location | Session |
|---|---|---|---|
| Mon, July 6 | Morning 8:50–12:25 | 北1-306 | Session 1 — Foundations (Lec 1–2) |
| Tue, July 7 | Afternoon 14:00–17:00 | 经济学院121报告厅 | Session 2 — ML Essentials (Lec 3) |
| Wed, July 8 | Morning 8:50–12:25 | 北1-306 | Session 3 — Deep Learning for Macro (Lec 4) |
| Thu, July 9 | Morning 9:00–12:00 | 经济学院121报告厅 | Session 4 — RL & Heterogeneous Agents (Lec 5–6) |
| Fri, July 10 | Morning 8:50–12:25 | 北1-306 | Session 5 — LLMs & Text (Lec 7–8) |
| Sun, July 12 | Morning 8:50–12:25 | 北1-306 | Session 6 — Agentic AI & Case Studies (Lec 9–10) |
Each of the six meetings taught one or two of the ten lectures. The Fall 2026 course keeps the same six-part arc but was rewritten from scratch as six sessions; this table maps the old meetings to the lectures whose decks are kept below.
| July 2026 meeting | Lectures | Closest Fall 2026 session |
|---|---|---|
| 1 — Foundations | 1–2 | Session 1 — Orientation |
| 2 — ML Essentials | 3 | Session 2 — Deep learning |
| 3 — Deep Learning for Macro | 4 | Sessions 2–3 |
| 4 — RL & Heterogeneous Agents | 5–6 | Sessions 3–4 |
| 5 — LLMs & Text | 7–8 | Sessions 5–6 |
| 6 — Agentic AI & Case Studies | 9–10 | Session 6 — RAG and agents |
The decks as they were posted for the Summer 2026 course, watermarked like the current ones. Decks for Lectures 6 and 10 were not posted. The notebooks that went with them are still on the labs page.
Three topic decks.
Two topic decks.
Four topic decks.
Four topic decks.
Three topic decks — one map, three acts: Model, Sample, Scale.
Five topic decks — measure, represent, attend.
Five topic decks — one map, five acts: the wall, index, retrieve, generate, apply. T1–T3 build the pipeline; T4 turns retrieved evidence into a cited answer; T5 puts it to work on real research corpora.
Eleven topic decks in five parts — see, under the hood, build, design & validate, run & govern — plus a dated appendix.
Students worked in teams of two to four on one of four tracks and submitted a reproducibility
package — paper, code repository, AI_LOG.md, a 12–15 minute talk, a
provenance package, and a referee report on another team's project — assessed on a 100-point
rubric with a bonus for a validated stretch finding. Tracks B, C and D drew on unpublished
research and carried confidentiality obligations for that cohort.
The current project tracks keep the same four projects, the same replicate-first rule, and the same validation checklists, but are written for self-directed work: the rubric became a self-assessment, the milestones became suggested pacing, and every track now runs on public data only.