Research Scientist/Engineer (Evaluations)

Research Scientist/Engineer (Science of Scheming)

Apollo Research

Application deadline: We are conducting interviews actively and aim to fill this role as soon as we find someone suitable.   ABOUT THE OPPORTUNITY   We want to develop a “Science of Scheming”. The goal is ambitious and we’re looking for Research Scientists and Research Engineers who are excited to build a new hard science from the ground up.   YOU WILL HAVE THE OPPORTUNITY TO   - Collaborate with leading AI developers. We partner with multiple labs, giving you access to a breadth of models that no single AI lab could offer. Through long-term research collaborations, your work directly impacts how the most capable AI systems are built and deployed. - Deeply study the RL dynamics that lead to the emergence of reward-seeking, evaluation awareness or misaligned preferences. Design and train model organisms, and scale your insights to frontier systems. - Work towards “Scaling laws of scheming”. Build the empirical foundations to predict how scheming risks evolve as models scale in capability. - Develop novel and ambitious evaluation techniques that have a chance of scaling to highly evaluation aware models. - Deep dive into AI cognition. Discover patterns in the reasoning processes of frontier AI systems that no one else has ever observed before.   Note:  We are not hiring for interpretability roles.   KEY REQUIREMENTS    A diverse range of skill sets will be required to drive our research agenda forward and we don’t expect any single candidate to fulfill all the characteristics below. That being said, a successful candidate likely displays excellence at one or several of the following:   - Fast-paced empirical research: You can design and execute experiments. You always strive to speed up iteration cycles and relentlessly drive progress towards the next empirical milestone. - Conceptual insights about scheming: You have deeply thought about the problem of AI scheming and are familiar with all the relevant literature. You are able to turn vague and undefined concepts into concrete and insightful experiment proposals. - Software engineering skills: Strong software engineering skills correlate highly with effective execution, even in an era of AI agents. Our entire stack uses Python. - Intense interest in AI progress: You always stay up to date on the latest model releases, and continuously tinker with new and creative AI workflows to speed up your work. You are fascinated by AI cognition and actively spend time trying to understand how they think. - Experience RL-training LLMs: You have hands-on experience in training LLMs via reinforcement learning. You have encountered and resolved countless painful issues from GPU failures to debugging learning instabilities. - Strong analytical skills: You bring rigorous quantitative chops from working on fields such as scaling laws in LLMs, statistical physics, dynamical systems, applied statistics etc. You're comfortable building mathematical models of empirical phenomena and know how to extract signal from noisy data.   We want to emphasize that people who feel they don’t fulfill all of these characteristics but think they would be a good fit for the position, nonetheless, are strongly encouraged to apply. We believe that excellent candidates can come from a variety of backgrounds and are excited to give you opportunities to shine. We don’t require a formal background or industry experience and welcome self-taught candidates.

Backend Software Engineer (Research team)

Apollo Research

Application deadline: We are conducting interviews actively and aim to fill this role as soon as we find someone suitable.    ABOUT THE OPPORTUNITY   We’re looking for Backend Software Engineers who are excited to build tools for frontier AGI safety research, e.g. building and maintaining evals libraries and tools for monitoring and controlling our own LLM traffic.   REPRESENTATIVE PROJECTS   Here is a list of example projects which you might build and ship in your first 6 months.   - Internal tooling for efficiently running and analyzing evaluations. For example, a tool that quickly investigates thousands of agentic eval runs in parallel and surfaces interesting information automatically - Automated evaluation pipelines to minimize the time from getting access to a new model for pre-deployment testing to analyzing the most important results and sharing them - Orchestration tools that allow researchers to run thousands of agentic evaluations in parallel on remote machines with high security and reliability - LLM proxy service that enables us to monitor all of our coding agent traffic in real time and identify undesired behavior automatically (in the spirit of Control) - LLM agents and MCP tools to automate internal software engineering and research tasks, with sandboxes to prevent major failures - CI pipeline optimisations to reduce execution time and eliminate flaky tests - Telemetry API and instrumentation of our existing tools, allowing us to monitor usage and improve reliability - Data warehousing pipeline and service to store thousands of eval transcripts which researchers can study and build datasets from - Upstream improvements to the Inspect framework and ecosystem, e.g. support for evaluating modern agentic scaffolds.

Full-stack Software Engineer (Research team)

Apollo Research

Application deadline: We are conducting interviews actively and aim to fill this role as soon as we find someone suitable.    ABOUT THE OPPORTUNITY   We’re looking for Full-stack Software Engineers who are excited to build tools for frontier AGI safety research, e.g. building and maintaining evals libraries and tools for monitoring and controlling our own LLM traffic.   REPRESENTATIVE PROJECTS   Your main objective is to develop tooling for analyzing model evaluation results. Here is a list of features that you might build and ship in your first 6 months:   - LLM-powered search that finds interesting fragments in evaluation transcripts - Comparison views that show how conversations and scores differ between two evaluation runs - Ability to view and analyse conversations with coding agents (Cursor, Claude Code, etc.) in addition to evaluation transcripts - Results streaming for evaluations that are currently being run - Collaborative editing of evaluation logs that automatically updates metrics and other derived data.   Think of this as developing an “IDE for evaluations”.   Besides this, here are example auxiliary projects which you might do:   - Automated evaluation pipelines to minimize the time from getting access to a new model for pre-deployment testing to analyzing the most important results and sharing them. - LLM agents and MCP tools to automate internal software engineering and research tasks, with sandboxes to prevent major failures - Telemetry API and instrumentation of our existing tools, allowing us to monitor usage and improve reliability - Upstream improvements to the Inspect framework and ecosystem, e.g. support for evaluating modern agentic scaffolds.