Research Engineer, Interpretability

Role overview Investigate failure modes in AI models that are approaching deployment, using mechanistic interpretability methods on the models themselves. You’ll turn suspicious behavior into reproducible tests, build tools that make findings independently checkable, and work closely with engineers whose deployment decisions may change as a result.

Responsibilities - Run interpretability experiments such as probing, sparse autoencoders, activation patching, and circuit tracing. - Convert isolated examples of unexpected behavior into reproducible tests for broader failure patterns. - Build scripts and notebooks that allow others to verify findings and compare activations against known patterns. - Document methods and results clearly for assurance engineers and client-facing teams. - Collaborate directly with engineers shipping the systems and communicate findings on short timelines.

Requirements - Hands-on experience with mechanistic interpretability methods, demonstrated through code. - Ability to inspect and modify model internals in PyTorch or JAX, including intermediate activations in models you did not train. - Skill in explaining technical results and uncertainty to people making deployment decisions. - Sound judgment in distinguishing real model behavior from artifacts introduced by an analysis method.

Nice to have - Publications or open-source contributions in interpretability are helpful but not required. - Experience finding real failure modes in deployed or production systems is also relevant.

Benefits and work setup - Remote research role.