Competitions

What our challenges look like

Each challenge fixes a target and a scoring harness, and asks entrants to improve the target with AI in the loop. Typical shapes:

• Use a model to write kernels that make the same model's training or inference faster.
• Improve an agent's own harness, scored on a held-out task suite.
• Build a data filter with a model, scored by the quality of the model trained on what it keeps.
• Find the cheapest mix of models that matches a stronger model on a fixed eval.

How a competition works

1. Pick a challenge and read the scoring rules.
2. Submit code. It is scored automatically on held-out evals.
3. Iterate in public on a live leaderboard.
4. Top entries are credited by name and, where a prize is attached, paid.

Elsewhere

Some of the best loop-relevant competitions run in other communities. Kernel leaderboards, where people and increasingly AI systems compete to write the fastest GPU code, are a good example. We track results like these on the benchmarks page, with credit to the organisers.

Propose or sponsor a challenge

Have a target, a harness and a metric? We can help turn it into a challenge.