About this role
About the Role Anthropic seeks Research Engineers to design and implement evaluations that quantify Claude's capabilities and behaviors. You will build scalable evaluation infrastructure and translate ambiguous notions of intelligence into defensible metrics, enabling researchers, leadership, and the public to understand model performance. You will collaborate across teams to define what to measure, run evaluations through live training checkpoints, and interpret results to advance safe, well characterized AI systems.
What You'll Do
- Design and run new evaluations of Claude's capabilities — reasoning, agentic behavior, knowledge, safety properties — and produce visualizations that make the results legible to researchers and decision-makers
- Build and harden the distributed eval execution platform so hundreds of evals run reliably against checkpoints throughout production RL training runs
- Own the dashboards researchers and leadership use to monitor model health during training, improving signal-to-noise, reducing latency, and making regressions impossible to miss
- Debug anomalous eval results mid-training-run, determine whether the cause is a model change or an infrastructure issue, and communicate the answer clearly under time pressure
- Improve the tooling, libraries, and workflows researchers use to implement and iterate on evaluations
- Partner with research teams across the full lifecycle of a new capability — from defining what to measure to interpreting results as training progresses
- Run experiments to characterize how prompting, sampling, and scaffolding choices affect results on internal and industry benchmarks
- Communicate evaluations and their results to internal stakeholders and, where appropriate, external audiences
What We're Looking For
- Strong Python programming skills, including production or research infrastructure
- Experience building or operating distributed systems, data pipelines, or other infrastructure that needs to be reliable at scale
- Clear written and verbal communication, especially when explaining technical results to non-specialists
- Comfort operating in an on-call or production-support capacity when training runs are live
- Care about the societal impacts of your work You are trained on data up to October 2023.