About the job
AI Research Scientist
Overview
Join Physical Superintelligence, an innovative startup rooted in prestigious institutions such as Harvard, MIT, Johns Hopkins, Oxford, the Institute for Advanced Study, and the Perimeter Institute. We are at the forefront of building AI systems designed to uncover groundbreaking insights in physics on a grand scale. We are in search of talented AI researchers dedicated to developing reinforcement learning agents and training frameworks that propel scientific discovery.
Key Responsibilities
- Develop and optimize AI systems aimed at physics discovery, collaborating with physicists on verification harnesses and engineers on training infrastructure.
- Address critical AI research questions related to agent learning in physics reasoning, action space design for scientific exploration, reward structure development, and scalable training systems.
- Construct and train reinforcement learning agents leveraging cutting-edge methodologies such as PPO, SAC, MuZero, and multi-agent self-play.
- Design agent architectures tailored for physics reasoning and scientific tool utilization.
- Execute training curricula and reward structures for discovery tasks.
- Establish evaluation workflows and benchmarks to assess physics reasoning capabilities.
- Develop instrumentation to analyze agent behavior and learning dynamics.
- Collaborate closely with physicists and engineers to refine system design and architecture.
Candidate Profile
We are looking for candidates with a strong background in developing agents and training models using reinforcement learning. Proficiency in modern machine learning frameworks and experience with distributed training systems is essential, alongside a proven track record of deploying effective AI systems.
Essential Skills:
- Practical experience with contemporary reinforcement learning algorithms including PPO, SAC, MuZero, and multi-agent self-play.
- Proficient in PyTorch or JAX, with hands-on experience in distributed training using Ray, XLA, or Accelerate, and familiarity with modern pretraining workflows.
Preferred Background:
- A strong foundation in physics or mathematics that enhances intuition for physical reasoning and mathematical modeling.
- Experience applying agents in simulators, games, scientific tool use, or benchmark design employing rigorous experimental methodologies.
