Open Research Seeds

Research ideas you can't be outscaled on.

Academic ML research is in a difficult moment. Lots of problems require scaling, billion-parameter models, massive RL runs, asking for compute that no university lab can match. It's hard for junior students to take their time on projects without the worry of getting scooped.

The ideas here are chosen with that pressure in mind. They are deliberately unconventional — grounded in mechanisms rarely seen in the scaling playbook. You can take your time with these. Each idea comes with a clear intuition, a concrete first experiment, risks, readings, and enough supporting material to get started.

These are open research questions that I myself found genuinely interesting but won't pursue myself. If you do choose to look at them and end up with a submission, no need to credit me — an acknowledgement to the Open Research Seeds effort would be enough. I hope you have fun.

Browse seeds Contribute an idea GitHub repo

For junior researchers

Pick a seed, read the entry point, and start with the smallest experiment. Each idea is meant to be open enough to become your own project.

For contributors

Senior students and researchers can add ideas through pull requests. Use the template so every seed has a clear hook, background, and first step.

For collaborators

Use GitHub Discussions to claim interest, ask questions, and coordinate with others before duplicating effort.

Seed Index

Choose a topic, then open a seed when something looks useful.

Selected Seed

Select a seed to open the full writeup.

Add a seed

Copy the template, write the idea card, add readings, and open a pull request. The goal is not guaranteed feasibility; it is a clear mechanism and a runnable first experiment.

  1. Copy ideas/_template/ to ideas/your-idea-name/.
  2. Fill in README.md and references.md.
  3. Add one entry to seeds.json.
  4. Open a pull request and start a Discussion thread if you want collaborators.