Interpretability Engineer, LLM Research & Infra

Anthropic; United States; Hybrid; USD 315,000 - 560,000; Full time

Benefits offered by this job

Remote work considered case-by-case

Visa sponsorship may be available

Office in San Francisco

Job summary

Anthropic is seeking an AI Engineer focused on interpretability and safety. The role involves building engineering systems that enable research into how large language models represent information and produce behavior.

You will work across model internals, distributed training and inference, accelerator performance, and researcher-facing tools to translate interpretability methods into dependable safety-audit workflows.

Qualifications

• 5–10+ years of professional software engineering experience.
• Proficiency in Python, Rust, Go or Java; ability to work productively in Python.
• Ability to investigate unfamiliar areas and trace bottlenecks across a system.
• Comfortable collaborating with researchers and engineers in a fast-moving environment.
• Genuine interest in interpretability and AI safety; prior interpretability experience not required.

Responsibilities

• Build and maintain specialized training and inference infrastructure for interpretability research.
• Support instrumented forward and backward passes, activation extraction, and controlled application of steering vectors.
• Profile systems, identify scaling constraints, and improve performance and efficiency across hardware and software layers.
• Create abstractions and platforms that let researchers run experiments quickly without unnecessary engineering friction.
• Help operationalize interpretability research in production safety audits with strong reliability expectations.
• Collaborate with infrastructure and research partners while working across the stack from model internals to user-facing tooling.

Skills

Python

Rust

Go

Java

Job description

Anthropic is seeking an AI Engineer focused on interpretability and safety. The role involves building engineering systems that enable research into how large language models represent information and produce behavior.

You will work across model internals, distributed training and inference, accelerator performance, and researcher-facing tools to translate interpretability methods into dependable safety-audit workflows.