How this course is made
FrontierCS 2 asks AI agents to do computer-science research: each task removes the key idea of a paper from the authors' own code, and an agent must write it back and beat the original. This course teaches those 48 tasks, one field at a time, so that a student can follow what each paper found and what the agents did.
What each lecture covers
For every task in a field, a lecture explains four things, in this order:
What the paper found
Built up from a small runnable example you can modify in your browser.
What the agent writes
The exact contribution cut from the authors' released repository.
How it is judged
The metrics, hidden workloads, and strict pass rules with their margins.
What the agents did
Who passed, the best scores against the reference, and the ideas behind the best runs.
Lecture 0 explains the benchmark itself: how a task is built, the pass rule, and how the leaderboard is computed.
Where the material comes from
Only public sources: the task statements an agent reads, the papers, and the task pages and case studies on frontier-cs.org. Nothing comes from a task's hidden workloads, its judge, or the authors' reference code, so the lectures do not give away how to pass a task. Small examples and their numbers are ours, and each lecture says so.
How each lecture is checked
- Every program runs. Each lecture starts as notes whose Python blocks are run by a checker; every output on the page is what the program really printed.
- Demos match the notes. Each interactive demo repeats a program from the notes and prints the same numbers.
- Results match the source. Pass counts, best scores and reference scores are compared with each task's page. Results change as models are added: every lecture states its date, and this edition uses the leaderboard as of October 26, 2026.
- Independent review. Before release, each lecture is read by a reviewer who did not write it: facts, technical accuracy, demos and clarity.
- Desktop and phone. Every page is checked at laptop and phone widths.
Source and contributions
The source of the course (notes, pages and checking tools) is public at
github.com/boxi-cs/frontiercs2-src.
Corrections are welcome as an issue or a pull request; its CONTRIBUTING.md explains how
to preview and check a lecture.
The benchmark
The tasks, results and case studies belong to the FrontierCS 2 benchmark. To run it yourself, see github.com/FrontierCS/FrontierCS-2.