Modeling: Multimodal AI (MAS.S60 / 6.S985)
Paul Liang, Dimitris Bertsimas, Jinhua Zhao & Sang-Gook Kim · MIT · Spring 2026freeSource ↗Schedule — slides and recordings per lecture ↗Syllabus (Google Doc) — grading, homework cadence, project timeline ↗Tutorial 1 · PyTorch introduction (Colab) ↗Tutorial 2 · Fine-tuning a code LLM (Colab) ↗Tutorial 3 · Multimodal LLM fine-tuning (Colab) ↗Course repo (MIT-MI/mmai-course) ↗
Added 2026-09-10 at Connor’s direction and deliberately OFF the ladder: multimodal modelling is not a rung on the token-to-task path, so a rung would be invented. Ungated because its stated prerequisite (6.390-level ML plus PyTorch) is not a unit here either. Its five graded homeworks and five reading assignments are Canvas-only; what is public and runnable is three Colab tutorials, twenty-four slide decks, fourteen recordings and the project templates — the scope below counts the lectures, not the homework.
No weekly slot allocated (Connor, 2026-09-10).
Environment
Setup is a real gatepartial~/dev/study/mit-mmai-spring2026/
The three tutorial notebooks are downloaded under tutorials/ with the proposal and midway LaTeX templates beside them. Nothing has been run: the notebooks target a Colab GPU, and no local run has been attempted.
Segments
The named subset — the scope above is the count| Ref | Title | Done | Artifacts | Note |
|---|---|---|---|---|
| T1 | Tutorial · PyTorch introduction | · | NBBLDGMEBTL0/0 | |
| T2 | Tutorial · Fine-tuning a code LLM | · | NBBLDGMEBTL0/0 | NEEDS — Colab GPU. |
| T3 | Tutorial · Multimodal LLM fine-tuning | · | NBBLDGMEBTL0/0 | NEEDS — Colab GPU. |
Artifacts
Consume → do → output- What fusing a second modality costs to serve — measured on a run tutorial, not readNot yet — nothing has landed for this slot.
Feeds
What closing this sharpensThe graded work (HW1–HW5, the reading assignments, the midterm) is not public, so this unit can never earn more than a `read it` seat on the lectures alone — the `ran it` seat has to come from the tutorials or a self-set project using the templates.