Application to Grants

At Grepped, we believe there is an abundance of talent. People who have the capacity to build RL environments that continue to maintain, and even accelerate, AI's vertiginous progress.

We also believe much of this talent is bottlenecked by data, compute and connections. To even begin pushing the post-training frontier, one needs access to beyond SOTA data, a lot of FLOPS and a trusted channel to the big labs. Most solo researchers don't have that.

Luckily, Grepped does.

Today we are announcing Grepped Grants. They are for researchers and small teams who turn real work into RL environments, evals and graders. If you're selected, you get three things.

Data. Access to real company archives that have never been on the open web: years of email, documents, tickets, code and records from operating businesses, de-identified before they reach you.
Compute. Credits for inference and training runs, so you can run baselines and show that training on your environment actually moves a model. Credits are released as you hit milestones.
Connections. A trusted channel to the post-training teams at the big labs. When your environment is ready, we take it to them, and you share in what it earns.

How we make money together. When a lab buys an environment you built, we first pay the company whose data it was built from their share. What is left is the profit, and we split it with you 50/50. You see what the lab paid, what went to the data owner, and how your half was worked out. If nothing sells, you owe us nothing.

You keep your code and tooling, and the environments stay non-exclusive. We agree a short, concrete plan with you, with clear milestones, and our team works alongside you as you build.

Grants are by application. Tell us what you've built, what you're working on, and what's holding you back.

Name Email X username (optional)

We need depth. Short or generic answers won't be considered. Be specific: name what you built, link it, and give numbers.

What have you built in this space? Name each environment or eval, link it, and say how many tasks it has. What are you currently working on? Where are you bottlenecked? Links to your work (optional): GitHub, papers, environments, evals Which one domain would you start with, and why would a lab pay for it? Show us a sample: 3 to 5 tasks with the grader and a reference run. Paste them, or link a repo or doc. How do frontier models score on your tasks today, and how does your grading resist gaming? Is it just you, or a team? Any affiliation? What compute do you need for a first version, and for what?

We read every application and reply by email.