Interpretability Researcher

Best AI Tools Wiki; San Francisco (CA); Hybrid; USD 180,000 - 230,000; Full time

Benefits offered by this job

Generous research budget

Conference and education stipend

Sabbatical program

Job summary

Anthropic is seeking an Interpretability Researcher to advance mechanistic interpretability, analyzing how neural networks represent information and exposing their internals. You will develop new techniques and contribute to transparent, trustworthy AI systems.

The role requires a PhD and strong mathematical foundations, with proficiency in Python and PyTorch/JAX, and effective communication of complex ideas across teams.

Qualifications

• PhD required in ML, neuroscience, physics, or mathematics.
• Strong mathematical foundations and research track record.
• Proficiency in Python and PyTorch or JAX.

Responsibilities

• Understand how neural networks represent and process information.
• Develop new techniques for mechanistic interpretability and analyze model internals.
• Contribute to making AI systems more transparent and trustworthy.

Skills

Python

PyTorch/JAX

Mathematics

Communication

Education

PhD in ML, neuroscience, physics, or mathematics

Tools

None

Job description

Anthropic is seeking an Interpretability Researcher to understand how neural networks represent and process information. You will develop new techniques for mechanistic interpretability, analyze model internals, and contribute to making AI systems more transparent and trustworthy.

Requirements

• PhD in ML, neuroscience, physics, or mathematics
• Research experience in interpretability or related fields
• Strong mathematical foundations
• Proficiency in Python and PyTorch/JAX
• Ability to communicate complex ideas clearly

Nice to Have

• Published work in mechanistic interpretability
• Experience with sparse autoencoders or probing methods
• Background in neuroscience or cognitive science
• Experience with visualization tools for neural networks

Benefits

• Generous research budget
• Conference and education stipend
• Sabbatical program

Skills

Interpretability Mechanistic Interpretability PyTorch Research Mathematics Python