Skip to main content
Computationally Assisted Metacognition
GitHub Toggle Dark/Light/Auto mode Toggle Dark/Light/Auto mode Toggle Dark/Light/Auto mode Back to homepage

About this course

Course
CIS 7000-008 · CRN 92792 · Fall 2026
Instructor
Jérémie Lumbroso, Practice Assistant Professor of Computer and Information Science — lumbroso@seas.upenn.edu
When
Thursdays, 5:15 PM–8:14 PM · first meeting Thursday, August 27
Where
Towne 217
How to enroll
Permit-only: 20 seats, by application. The application link is being finalized — write to the instructor in the meantime.

Computationally Assisted Metacognition (CAM) — thinking better, with AI, across domains. A seminar on using frontier language models as instruments of reflection: students iterate live with models, read their reasoning traces, and turn their own thinking into something they can inspect, critique, and revise. Sessions are hands-on and discussion-driven, built around exercises drawn from real human–AI collaborations. A technical spine for CIS majors and graduate students (a conceptual track for others): how frontier models are built and run; running local models; API calls against local and cloud models; building and deploying MCP servers and tools; how context is composed and retrieved. Culminates in a build — a harnessed personal agent each student constructs and deploys. Open to graduate and undergraduate students by application; waitlist.

What you should be able to do by the end

By the end of the course, students can:

  1. Use a frontier model as an instrument of reflection — externalize a question, iterate, and read the model’s reasoning trace critically rather than accept its answer.
  2. Recognize and name metacognitive moves — reframing, decomposition, steelmanning, falsification — in their own conversation transcripts.
  3. Explain how frontier models are built and run (architecture, training, inference, context) well enough to predict their failure modes.
  4. Build and deploy tools a model can call — API clients, MCP servers — locally and remotely.
  5. Construct, deploy, and document a harnessed personal agent, including the design decisions behind it.

How the room works

Sessions use Slido throughout, with names attached. In a course on metacognition we are invested in owning our own opinions: taking a stance, being wrong — including being wrong in public — is not a bad thing. What we are developing is the practice of adapting to new information and new evidence.