Research Scientist, Takeoff Intel

Anthropic•San Francisco, CA

20d•$350,000 - $850,000•Hybrid

About The Position

Anthropic is seeking a Research Scientist to focus on measuring and understanding recursive self-improvement in AI systems. The ideal candidate will have hands-on research experience with large models (pretraining, fine-tuning, RL, evals, or agent scaffolds) and a deep understanding of the model development loop. This role involves identifying key signals for AI R&D acceleration, designing evaluations, building quantitative models of capability growth, running experiments, and making research bets. The position is available at both junior and senior levels, with senior researchers expected to contribute to setting research direction alongside technical work. The role emphasizes impact, clear communication, and a commitment to AI safety.

Requirements

Hands-on research on large language models: pretraining, fine-tuning, RL, evals, or agent systems
Strong quantitative instincts, comfortable with quantitative modeling and reasoning
Experience in forecasting, may have published AI forecasting scenarios
Ability to design an evaluation from a vague question and defend the methodology
Clear writing and calibration: state confidence, name what would change your conclusion
Motivated by impact: comfortable with work whose output is graded assessments and system-card sections more often than papers
Care about AI safety and think carefully about where rapid capability growth leads

Nice To Haves

Trained or RL'd frontier models hands-on
Experience with scaling laws, capability forecasting, or emergent-capability studies
A physics, applied-math, or similarly quantitative background that moved into ML
Written a system card section, capability report, or methodology document that others cite
Experience supervising and correcting AI-written code

Responsibilities

Identify the signals that track AI R&D acceleration and design the evaluations that measure them
Build quantitative models of capability growth and self-improvement dynamics, grounded in evaluation and telemetry data
Run experiments and evals to test hypotheses about automation and capability
Make opinionated research bets and own the outcome
Write graded assessments of what our measurements show, for internal decision-makers and public reporting
Collaborate with pretraining, RL, economic research, and policy teams