Research Scientist/Engineer (Evaluations)
Apollo Research8 months ago
London, United Kingdom or San Francisco, CA, USASenior
Responsibilities
- Run and own pre-deployment evaluation campaigns for frontier AI labs, analyzing model behavior across thousands of runs and hundreds of environments.
- Evaluate and red-team model checkpoints throughout post-training to identify alignment faking, scheming, and other undesirable behaviors.
- Design and improve training-run assessment methodology, including implementing new evaluations and automated red-teaming infrastructure.
- Build and maintain efficient Python pipelines for automated analysis of post-training data.
- Conduct targeted follow-up experiments and communicate qualitative and quantitative findings to technical and non-technical audiences.
Requirements
- Strong software engineering experience, ideally including shipping and maintaining production Python code and creating clean, extensible abstractions.
- Ability to extract signal from large, messy datasets, perform quantitative analysis, identify anomalies, and determine when qualitative assessment is appropriate.
- Strong written and verbal communication skills for conveying technical and non-technical findings concisely.
- Ability to use AI tools effectively and evaluate which models and workflows are appropriate for different tasks.
- Formal education and industry experience are not required; self-taught candidates are welcome.
- Preferred: experience with AI risk topics such as scheming and metagaming.
- Preferred: knowledge of LLM post-training, including RLHF, reasoning training, supervised fine-tuning, and on-policy distillation.
- Preferred: experience creating evaluations with Inspect or similar frameworks such as Harbor.
Benefits
- Market-competitive salary, equity, and competitive benefits.
- Flexible work hours and schedule.
- Unlimited vacation and sick leave.
- Up to six months of paid parental leave.
- Comprehensive health, dental, and vision insurance.
- Retirement savings with competitive employer matching, including 401(k) for US employees.
- Lunch, dinner, and snacks provided on workdays.
- Paid work trips, staff retreats, business trips, and relevant conferences.
- A yearly $1,000 USD professional development budget.
- Relocation support and visa fees when applicable.
- Full-time, in-person work from the London or San Francisco office, with flexible hours and work-from-home arrangements.
Tech Stack
Categories
AI Research
About Apollo Research
Apollo Research evaluates and stress-tests frontier AI models for safety risks and builds security tools for agentic systems used by enterprise engineering teams. It works with OpenAI, Anthropic, Google, Meta and others to run pre-deployment tests focused on deceptive behavior and loss-of-control risks, and sells Watcher, a coding-agent security product that monitors billions of tokens in production. Revenue comes from enterprise software and research partnerships with AI labs and agent-building companies.