Researcher, Evaluations and Benchmarks

This is a fully remote position, open to applicants in California.

📋 Description

• Deliver a benchmark every two to three weeks that evaluates a previously unmeasured frontier risk.

• Partner with top AI laboratories and universities to work on benchmarks and academic papers.

• Manage the benchmark taxonomy, harness, quality standards, and release processes.

• Personally review evaluations and ensure alignment with taxonomy, rubrics, verifiers, and data distribution.

• Oversee the benchmark plan and schedule, ensuring researchers adhere to timelines.

• Supervise two or three freelancers/subject matter experts as required.

• Meet approximately monthly with the CTO, pod, and research leaders to update the quarterly release plan.

• Connect the release roadmap to targeted accounts.

• Dedicate roughly 20% of time to monitoring the ecosystem and reviewing research.

• Maintain relationships within AI labs and engage with lab personnel weekly.

• Travel to conferences a few times each year.

⛳️ Requirements

• PhD or Master's degree in computer science, machine learning, or a related area, or equivalent experience from industry research.

• Over 3 years of experience in building and executing safety or security evaluations for language models in production, within an AI lab, a model provider, or a safety and security research organization.

• At least 5 relevant research publications in AI safety and security, including being the lead author on a minimum of 2.

• Proficient engineering skills, encompassing evaluation harnesses, distributed inference, vLLM, and the ability to read and modify codebases.

• Capability to construct a taxonomy, not just evaluate against one.

• Proficiency in guiding a researcher and two freelancers without formal management responsibilities.

• Excellent command of English, both written and spoken.

• Eagerness to learn about AI harms and adapt to new subjects every three weeks.

• Ideally: experience post-training with SFT, DPO, or GRPO.

• Ideally: experience with agentic evaluation that includes tool use, orchestration, permissions, and prompt injection.

• Ideally: publications in leading conferences.

• Willingness to present personal work during client calls.

• Strong verbal and written communication skills, with the ability to present to large and/or senior audiences.

• Openness to travel to conferences at least 3 times a year.

🏝️ Benefits

• Budget for engaging freelancers (subject matter experts) as needed.

• Travel to conferences at least 3 times annually.

• Opportunity to collaborate with prestigious AI labs and universities.

• Access to around 150 researchers focused on AI harms.

• Conference travel a few times each year / at least 3 times a year.