Research Engineer, Interpretability

Anthropic's Interpretability team is seeking a Research Engineer to build specialized inference and training infrastructure for mechanistic interpretability research. The role involves instrumenting model internals, resolving scaling bottlenecks, designing tools for researchers, and bringing research into production safety audits. The work is akin to reverse-engineering neural networks to understand how they work, with a focus on AI safety and reliability.

5-10+ years of software engineering experience, proficiency in Python and at least one other language (e.g., Rust, Go, Java), strong ability to prioritize and collaborate, curiosity about AI interpretability and safety, comfort working with researchers. Preferred: experience with large-scale distributed systems, LLM optimization, PyTorch/CUDA or JAX/XLA, and building research tooling.