Senior Interpretability Systems Engineer for Safe AI

Anthropic Limited; San Francisco (CA); Hybrid; USD 315,000 - 560,000; Full time

Job summary

Anthropic Limited in San Francisco seeks a seasoned software engineer to join the Interpretability team, building and maintaining the infrastructure that powers interpretability research. You will instrument forward/backward passes, extract activations, and apply steering vectors to models.

You will work with researchers to scale and optimize pipelines, design abstractions, and push research into production safety audits with robust reliability.

Qualifications

• 5+ years of experience building software.
• Proficient in at least one programming language (e.g., Python, Rust, Go, Java) and productive with Python.
• Curious about interpretability research and its role in AI safety (though no research experience required).
• Strong ability to prioritize impactful work and operate with ambiguity and assumptions.
• Prefers fast-moving collaborative projects over solo work.

Responsibilities

• Build and maintain inference and training infrastructure powering interpretability research.
• Resolve scaling bottlenecks through profiling and optimization with infra teams.
• Design tools and platforms enabling researchers to experiment rapidly.
• Support production safety audits with high reliability requirements.
• Collaborate across stack from model internals to research tooling.

Skills

Software development

Python

Rust

Go

Java

Curiosity

Education

Bachelor's degree or higher

Relevant field coursework

Tools

PyTorch

CUDA

JAX

Job description

Anthropic Limited in San Francisco seeks a seasoned software engineer to join the Interpretability team, building and maintaining the infrastructure that powers interpretability research. You will instrument forward/backward passes, extract activations, and apply steering vectors to models.

You will work with researchers to scale and optimize pipelines, design abstractions, and push research into production safety audits with robust reliability.