Research Engineer, Safety Reasoning

About the Team

The team is at the forefront of OpenAI's mission to build and deploy safe AGI, driving our commitment to AI safety and fostering a culture of trust and transparency.

The Safety Reasoning Research team is poised at the intersection of short-term pragmatic projects and long-term fundamental research, prioritizing rapid system development while maintaining technical robustness.

Key focus areas include improving foundational models’ ability to accurately reason about safety, values, and questions of cultural norms, refining moderation models, driving rapid policy improvements, and addressing critical societal challenges like election misinformation.

As we venture into , the team seeks talents adept in novel abuse discovery and policy iteration, aligning with our high-priority goals of multimodal moderation and ensuring digital safety.

About the Role

The role involves developing innovative machine learning techniques that push the limit of our foundation model’s safety understanding and capability.

You will engage in defining and developing realistic and impactful safety tasks that, once improved, can be integrated into OpenAI's safety systems or benefit other safety / alignment research initiatives.

Examples of safety initiatives include moderation policy enforcement, policy development using democratic input, and safety reward modeling.

You will be experimenting with a wide range of research techniques not limited to reasoning, architecture, data, and multimodal.

In this role, you will :

Conduct applied research to improve the ability of foundational models to accurately reason about questions of human values, morals, ethics, and cultural norms, and apply these improved models to practical safety challenges.

Develop and refine AI moderation models to detect and mitigate known and emerging patterns of AI misuse and abuse.

Work with policy researchers to adapt and iterate on our content policies to ensure effective prevention of harmful behavior.

Contribute to research on multimodal content analysis to enhance our moderation capabilities.

Develop and improve pipelines for automated data labeling and augmentation, model training, evaluation and deployment, including active learning process, routines for calibration and validation data refresh etc.

Experiment and design an effective red-teaming pipeline to examine the robustness of our harm prevention systems and identify areas for future improvement.

You might thrive in this role if you :

Are excited about of building safe, universally beneficial AGI and are aligned with

Possess 5+ years of research engineering experience and proficiency in Python or similar languages.

Thrive in environments involving large-scale AI systems and multimodal datasets (a plus).

Exhibit proficiency in the field of AI safety, focusing on topics like RLHF, adversarial training, robustness, fairness & biases, which is extremely advantageous.

Show enthusiasm for AI safety and dedication to enhancing the safety of cutting-edge AI models for real-world use.