Member of Technical Staff — ML Research, Interpretability

We look for domain experts who are excited to tackle unsolved problems. Our thesis is that scaling on physics yields a model capable of understanding the causal structure to predict and alter the future. Your mission is to ensure the model evolves towards this thesis: grounded in physical law, evaluated against it, and ready to generalize across domains.

We look for researchers who are excited to tackle unsolved problems. Our mission is to build models capable of learning the underlying causal structures of physical systems and interpretability is how we will know whether they have. Before anyone acts on a model's prediction or its recommended intervention, we need to understand what the model has actually learned. Your mission is to open the model up: to understand its internal representations, explain its outputs, and build the trust that acting on physical systems demands.

Responsibilities

- Probe the model's internal representations for physical quantities, structure, and conservation laws

- Develop methods to explain individual predictions and the model's reasoning about interventions

- Investigate whether interventions in the model's internal state produce physically coherent responses

- Build tools and techniques for debugging model failures and understanding rollout behavior

- Partner with model, evaluation, and domain teams to turn interpretability findings into better models and greater trust

What we're looking for

We value a relentless approach to problem-solving, rapid execution, and the ability to quickly learn in unfamiliar domains.

- Strong grasp of machine learning fundamentals and the internals of modern neural network architectures

- Experience or strong interest in interpretability, representation analysis, or related research

- Strong engineering skills for building interpretability tooling and running careful experiments

- A rigorous, hypothesis-driven approach to understanding model behavior

- A track record of turning open-ended research questions into concrete findings