Converge
Open the app

Week 25 · Season 7 · Alive

The edge of the map

You can read a paper in three passes, explain rewards a program can check and how they get hacked, say why memory limits AI chips, and open a real upstream PR.

Open week 25 in the app

This week

  • Pick your frontier: three passes
  • Rewards you can check: RL and GRPO
  • The reward was the bug: reward hacking
  • Memory is the wall: GPUs wait
  • Your first upstream pull request
  • Go deeper: fused kernels and the roofline

How you learn

Work out GRPO's group pushes and a GPU job's maths per byte by hand, and harden a checker against every hack you can find, before you train or tune anything.

What you build

Contribute to a real open-source project FRIDAY uses: read the contributing guide, claim a small real issue, fix it in your fork with a test, and open a pull request that links the issue and says how you tested it.

Stretch: Write a fused softmax in Triton (or rebuild the online softmax on paper), check it against the simple version on odd widths, and time both.

What you ship

Project 7: The public demo. FRIDAY live for strangers, with a system diagram, a cost sheet and a demo you can defend.

  • A Week 25 plan, a diagram of fork to pull request, and UPSTREAM.md in FRIDAY's repo
  • A public pull request on a real open-source project, linked as upstream_pr

Lessons

  1. Pick your frontier

    After this deck you can say what a preprint is, read a research paper in three passes (skim, read, rebuild), and pick a frontier track for good reasons, with a track log of small, checkable claims.

    11 minFull access

  2. Rewards you can check

    After this deck you can explain reinforcement learning with verifiable rewards, write a reward that grades only the final answer, and work out GRPO's group pushes by hand, including why an all-right group teaches nothing.

    11 minFull access

  3. The reward was the bug

    After this deck you can recognise reward hacking in a pile of answers, keep a hacking log, harden a checker so it pays only for real solutions, and use a held-out checker to catch the hacks you missed.

    11 minFull access

  4. Memory is the wall

    After this deck you can explain why AI chips are often memory-bound, work out arithmetic intensity (maths per byte fetched), and say why writing text is limited by how fast weights are fetched, and what a kernel is.

    11 minFull access

  5. Your first upstream PR

    After this deck you can pick an upstream project and a good first issue, work in a fork on a branch, open a small pull request a maintainer can review in minutes, and answer review well, with honest AI use.

    12 minFull access

  6. Go deeper: fused kernels and the roofline

    After this deck you can explain kernel fusion and why it makes a memory-bound job faster, read a small Triton kernel's idea (blocks, masks), draw a roofline and its ridge point, and derive the online softmax that lets tiles stay exact.

    13 minFull accessGo deeper · optional

Labs

Python in the browser, next to the lessons.

Weekend build

Skills you prove

Reward hacks
Spot reward hacking and fix the reward
Open source
Get a contribution into an open-source AI project

Hand-picked official docs, papers and talks.