Research Scientist/Engineer (Evaluations)

We’re looking for Research Scientists/Engineers for our pre-deployment team to work on Training-Run Assessments (TRAs). You will design and build automated pipelines for assessing whether egregious misalignment or scheming are emerging at any point of frontier post-training. 

This will involve evaluating and red-teaming of checkpoints at various stages of post-training as well as automated analysis of post-training data.

You will get to work with frontier labs like OpenAI, Anthropic, and Google DeepMind and be among the first to interact with new models before anyone else. Our ideal candidate loves rigorously testing frontier AI models, and enjoys building efficient pipelines for automated analysis. 

• Run and own pre-deployment engagements: We run a pre-deployment evaluation campaign with a frontier AI lab approximately every two weeks with thousands of runs across hundreds of distinct environments. We explore behaviors learned during training and check for undesirable behaviors like alignment faking, perform targeted follow-up experiments/red-teaming, and report our findings to the frontier AI lab we’re working with.
• Develop methodology for training-run assessments: in-between campaigns we improve our methodology, which might mean implementing new evals or building infrastructure for automated red-teaming.

• We don’t require a formal background or industry experience and welcome self-taught candidates.
• Software engineering skills: Our entire stack uses Python. We're looking for candidates with strong software engineering experience. Ideally, you have experience shipping and maintaining production Python code, and know how to factor messy problems into clean abstractions that others can use and extend.
• Data Analysis & Pattern Recognition: You can extract signal from large, messy datasets. You're comfortable with quantitative analysis and know when qualitative assessment is more appropriate. You can identify anomalies and unexpected model behaviors.
• Writing and communication: You succinctly convey qualitative and quantitative findings to a technical and non-technical audience.
• AI power-user: You’re capable of using AI to accelerate your work, technical or otherwise. You have experience using different models, know which ones to use for which tasks, when not to use AI, and always experiment with new AI workflows.

• Experience thinking about AI risk topics like scheming and metagaming.
• Knowledge of LLM post-training: topics like RLHF, reasoning training, supervised fine-tuning, on-policy distillation, etc.
• We are using Inspect as our primary evals framework, and we value experience creating evals with it or similar frameworks like Harbor.