ABOUT ABUNDANT

Abundant is an applied research lab focused on scaling reinforcement learning for safe and reliable agentic capabilities. We are an extremely talent-dense team of researchers, roboticists, founders, and operators whose work includes the Waymo Driver.

 

THE ROLE

You will own a research agenda where research meets production, co-designing strategies with researchers from frontier AI labs. You are the “PM of the model”: you pick the problem, design the data and reward signal that moves it, and see it through to a live training run. The job rewards a founder mentality and extreme ownership, including going into unfamiliar domains, from chemical engineering to complex legal logic, and getting to ground truth fast on problems others have written off.

 

WHAT YOU’LL DO

• Foundational capability research and execution. Architect and execute a research agenda to find simple, generalizable ideas that advance model reasoning at scale, and own it from hypothesis through to running in live systems.

• Design tasks and verifiers that stay hard. Build long-horizon agentic tasks that are still difficult a model generation from now, with graders where partial credit means something and the obvious exploit does not score.

• Diagnose failure modes. Read trajectories until you can say why a model fails a class of work, then turn that into the data or reward signal that fixes it. This is the core loop of the job.

• Model alignment and data strategy. Partner with the world’s most advanced AI research teams on datasets and benchmarks that shape how frontier models behave, focused on alignment, safety, and reward signal design.

• Autonomous problem selection. Identify, scope, and run long projects, choosing the problems that matter most for scaling data toward AGI.

• System infrastructure. Work with engineering on data pipelines, internal tooling, and deep learning implementations. If a workflow produces bad data, you fix the workflow.

WHO YOU ARE

• You write a lot of code. Strong software engineering and deep Python, comfortable owning a system end to end and debugging through layers you did not write.

• You have built evaluations, RL environments, agent harnesses, verifiers, or execution sandboxes, and you have opinions about which of them can be trusted.