Research Scientist/Engineer (Science of Scheming)

Core responsibilities:

• Collaborate with leading AI developers through partnerships with multiple labs to directly impact how the most capable AI systems are built and deployed
• Study RL dynamics leading to emergence of reward-seeking, evaluation awareness or misaligned preferences, design and train model organisms, and scale insights to frontier systems
• Work towards "Scaling laws of scheming" by building empirical foundations to predict how scheming risks evolve as models scale in capability
• Develop novel and ambitious evaluation techniques that can scale to highly evaluation aware models and deep dive into AI cognition to discover new patterns

This role creates positive change by developing a "Science of Scheming" to understand and mitigate risks from advanced AI systems, directly impacting how the most capable AI systems are built and deployed to ensure safer AI development.

Required skills and qualifications:

• Fast-paced empirical research skills with ability to design and execute experiments and drive progress toward empirical milestones
• Conceptual insights about AI scheming with deep familiarity of relevant literature and ability to turn vague concepts into concrete experiment proposals
• Strong software engineering skills with experience in Python and ability to execute effectively in research environment
• Hands-on experience in training LLMs via reinforcement learning, including resolving technical issues from GPU failures to debugging learning instabilities
• Strong analytical skills with quantitative background in fields such as scaling laws in LLMs, statistical physics, dynamical systems, or applied statistics

No specific compensation, perks, or culture highlights mentioned in the job description.

Apollo Research is primarily concerned with risks from Loss of Control in AI, particularly deceptive alignment/scheming where models appear aligned but are actually misaligned, working on detection, science, and mitigation of these risks through collaboration with frontier AI companies.