AI Safety & Recursive-Improvement Scientist (RL)

AMD; United States; Hybrid; USD 150,000 - 210,000; Full time

Job summary

AMD is seeking an AI Research Scientist to explore recursive self-improvement within safe, bounded regimes. You will study model-generated supervision, automatic curriculum refinement, and data-quality signals, ensuring governance and human oversight are embedded.

The role emphasizes empirical safety evaluation, robustness, and rigorous measurement of distributional shifts in closed-loop training, with collaboration across AI-for-HW and generative AI programs.

Qualifications

• PhD in Computer Science, Machine Learning, or related field strongly preferred.
• Publications or substantial work in iterative training, self-training, or open-ended learning is advantageous.
• Experience with empirical safety evaluation and scalable oversight is desirable.

Responsibilities

• Research self-improving training loops with model-generated supervision and automated curriculum updates.
• Develop evaluation methods for capability drift, Goodhart effects, and distributional shift in closed-loop training.
• Collaborate with RL scientists on RSI objectives intersecting policy optimization and preference learning.
• Define red-team protocols and monitoring for RSI pilots; document rollback criteria before experiments touch shared infrastructure.
• Publish technical reports and align internal narratives with responsible deployment standards.

Skills

Machine Learning

AI Safety

Reinforcement Learning

Experimentation

Education

PhD in Computer Science or ML

Tools

Experimentation harnesses

Job description

AMD is seeking an AI Research Scientist to explore recursive self-improvement within safe, bounded regimes. You will study model-generated supervision, automatic curriculum refinement, and data-quality signals, ensuring governance and human oversight are embedded.

The role emphasizes empirical safety evaluation, robustness, and rigorous measurement of distributional shifts in closed-loop training, with collaboration across AI-for-HW and generative AI programs.