Research Scientist, Interpretability
$350k - $850k • Remote • San Francisco, CA
Posted 10d ago
Job Location
San Francisco, CA
Tech Stack
Remote Work Policy
Fully remote
Categories
AI Research Engineer
About the job
The Interpretability team at Anthropic is dedicated to reverse-engineering how trained models work, believing that a mechanistic understanding is crucial for making advanced AI systems safe and steerable. This role focuses on mechanistic interpretability, aiming to discover how neural network parameters map to meaningful algorithms, akin to performing "biology" or "neuroscience" on neural networks. The team has made significant progress in understanding issues like "superposition" and decomposing models into interpretable components, with recent work identifying millions of features in production models and developing methods to understand multi-hop reasoning and planning. This research is foundational for building reliable, interpretable, and steerable AI systems and often involves collaboration with other teams at Anthropic to enhance AI safety.
Responsibilities
- Reverse-engineer how trained neural networks work to achieve a mechanistic understanding.
- Discover how neural network parameters map to meaningful algorithms.
- Develop methods to decompose models into more interpretable components.
- Address issues like "superposition" that hinder model interpretability.
- Identify and analyze features within large language models.
- Develop and apply methods to understand complex model computations like multi-hop reasoning and planning.
- Collaborate with other teams (e.g., Alignment Science, Societal Impacts) to use interpretability research for improving AI safety.
Requirements
- Experience or strong interest in mechanistic interpretability of neural networks.
- Ability to develop novel methods for understanding AI systems.
- Proficiency in relevant programming languages and ML frameworks.
- Strong research and analytical skills.
- Ability to collaborate effectively with a research team.