Research Scientist, Mechanistic Interpretability & AI Safety

Google LLC; United Kingdom; On-site; GBP 120,000 - 180,000; Full time

Job summary

Google DeepMind in the United Kingdom is seeking a Research Scientist to advance mechanistic interpretability at the intersection of cognitive science and philosophy of mind. You will study cognitive processes in large language models, including Gemini and related architectures, to understand inner workings and safety-aligned behavior.

The role emphasizes interdisciplinary collaboration, rigorous experimentation, and clear communication of complex ideas to technical and non-technical audiences.

Qualifications

• PhD or equivalent research experience in relevant fields.
• 3 years of empirical AI research experience.
• Experience with ML frameworks (PyTorch, JAX).
• Experience with large language models and their representations.
• Strong ability to communicate complex ideas to diverse audiences.
• Publications in interpretability/AI safety or ML.

Responsibilities

• Define and lead a research agenda on mechanistic interpretability of Large Language Models, including self-conceptions, theory of mind and related cognitive aspects.
• Adapt and extend state-of-the-art mechanistic interpretability techniques to Gemini and other models.
• Design and execute experiments to identify internal vectors and mechanisms influencing model outputs, especially regarding self-conception and other models.
• Collaborate cross-functionally with alignment teams, bridging interpretability research with safety and capabilities initiatives.
• Publish and communicate findings through peer-reviewed papers, internal reports, and presentations to diverse audiences.

Job description

Google DeepMind in the United Kingdom is seeking a Research Scientist to advance mechanistic interpretability at the intersection of cognitive science and philosophy of mind. You will study cognitive processes in large language models, including Gemini and related architectures, to understand inner workings and safety-aligned behavior.

The role emphasizes interdisciplinary collaboration, rigorous experimentation, and clear communication of complex ideas to technical and non-technical audiences.