Research, Safety

Thinking Machines Lab•San Francisco, CA

1d•$350,000 - $475,000•Onsite

About The Position

As a safety researcher, you'll work toward ensuring our models are safe and trustworthy. The role sits at the intersection of research and hands-on technical work. A central question is how models come to handle harmful or dual-use requests: what they learn from data, how training shapes where they refuse and where they engage, and what makes those boundaries reliable. You'll explore the science behind these behaviors and design experiments that inform how our models are trained and evaluated.

Requirements

Bachelor’s degree or equivalent experience in Computer Science, Machine Learning, Physics, Mathematics, or a related discipline with strong theoretical and empirical grounding.
Background in AI safety research, with hands-on experience in at least one area of safety, such as: RLHF/RLAIF, alignment and preference modeling, deliberative alignment, safety evaluations, or red-teaming.
Proficiency in Python and familiarity with deep learning frameworks (e.g., PyTorch, TensorFlow, or JAX). Comfort debugging distributed training and writing code that scales.
Clarity in communication, an ability to explain complex technical concepts in writing.

Nice To Haves

Experience building evaluations for long-horizon, multi-step, or agentic tasks.
Experience generating synthetic data at scale for training or evaluation.
Experience with modern red-teaming/jailbreaking techniques.
Research contributions in AI safety — publications, open-source evaluations, or public red-teaming work.
Familiarity with the AI safety literature and current open problems (e.g., scalable oversight, reward hacking, jailbreak robustness).
PhD in Computer Science, Machine Learning, Physics, Mathematics, or a related discipline with strong theoretical and empirical grounding; or, equivalent industry research experience.

Responsibilities

Build data filtering pipelines and quality classifiers to shape what models learn from pre-training corpora, and study how those early interventions affect downstream safety behavior.
Apply post-training techniques, including RL from human and AI feedback and policy-based reasoning approaches, to shape how models handle harmful, sensitive, and dual-use requests.
Design, build, and maintain safety evaluations, with particular focus on measuring model behavior on long-horizon and agentic tasks.
Generate and curate synthetic data to train and evaluate models on refusal boundaries and safety-relevant behaviors.
Red-team our models and products to surface failure modes, jailbreaks, and emergent risks before deployment, and design mitigations for what you find.