Researcher, Pretraining Safety
Core
Developing upstream safety evaluations and safe-by-design architectures to ensure base models are safe before deployment.
Role type
Researcher, Pretraining Safety
Builds
Safer base models and pretraining pipelines for OpenAI's AGI mission
Domain
AI Safety / Large Language Model Pretraining
Deliverable
production ML models
Required skills
Pretraining architecture design, training infrastructure management, experimental design, statistical reasoning, data curation, loss function design, evaluation framework development
Preferred skills
Experience with diffusion models, multimodal models, Python, PyTorch/JAX, Apache Beam
Technologies
Python, PyTorch, JAX, Apache Beam
Responsibilities
Identify safety-relevant behaviors in early-stage models, design data curation strategies to improve pretraining priors, introduce novel safety-oriented loss functions and metrics, collaborate with cross-functional safety teams to unify risk reduction
Seniority
Mid-Senior, hands-on IC