Grants

What we fund

• Loop experiments: a coding agent improving its own harness, a model curating its successor's data, an AI-written kernel used back in training. The smallest version that shows the loop turning once is a great first grant.
• Open tools that make loops easier to run: eval harnesses, sandboxes, kernel test benches, experiment trackers.
• Benchmarks and datasets that measure AI doing AI research and engineering.
• Replications and audits of published self-improvement results, including the ones that do not hold up.
• Compute, API credits, data and annotation costs for any of the above.
• Paid contributor time on a loop project that is already moving.

What we do not fund

• A strong tool or model with no stated path back into AI development.
• Capability work whose only lever is more compute or more data.
• Closed research with no public output.
• Salaries for full-time staff of frontier AI labs, or tuition for degree programs.

What a strong application says

1. The loop, in one sentence. What the AI does, what it improves, and how that improvement comes back.
2. How you will know it turned. The measurement, and the baseline you compare against.
3. The kill test. The smallest result that would change your mind, and what it costs to find out.
4. What you will publish. Code, data, a write-up, or all three.

Sizes and timing

Grants are sized to the work. Most cover a focused experiment or a few months of part-time effort: compute and API credits, data, or paid time. Larger grants are possible for projects that have already shown their loop turning once.

There are no fixed rounds or deadlines. We read applications as they arrive.

How decisions are made

Each application is read by reviewers who work on related problems. We do not ask for a CV: we judge the proposal and any public work you link. Decisions usually take a few weeks, and if we say no, we tell you why.