Today, we are announcing Proximal.

The biggest bottleneck to superhuman coding agents is data. Nobody knows what the best training data looks like to create models 10-100x better than today. As models improve, they'll generate more of their own training data. Making this work requires serious research and engineering.

We're building a data research lab to solve this. We experiment constantly to figure out what training data pushes capabilities forward. When we find something that works, we build systems to generate it at massive scale. Our goal is simple: build the infrastructure that lets models create their own high-quality training data autonomously.

Our founding team has worked on RL infrastructure and coding agents at companies like Cursor and Prime Intellect and has previously founded and successfully sold companies. We are backed by top investors and angels from OpenAI, Anthropic, Thinking Machines, xAI, Meta Superintelligence, Cursor and Cognition.