Task Development Engineer

Team: Engineering & Research

Location: Berkeley

Commitment: Employee

Workplace Type: onsite

Salary:

About the role

• As part of informing the world about risk from frontier AI systems, METR often runs and publishes evaluations of frontier models.

• Time Horizons is a central tool the world uses to understand AI progress. Our methodology has been included in system cards, called an "obsession" by the NYT, has wide reach online, and is used by governments to inform national policy. It is essential to our broader risk assessment work to have good capability evaluations.

• Task Development Engineers contribute to METR’s expanding ambition of our evaluations with high quality tasks supporting the Time Horizons methodology. We expect our results to be seen by policymakers, frontier labs, national security stakeholders, and other key decisionmakers influencing society’s response to AI progress.

Note: We recently changed this role to be a full time role by default (though we are happy to discuss contractor setups if you prefer that).

What this role looks like

• (Primarily, and most importantly) Developing difficult, novel tasks for models. You will build well-scoped tasks that remain challenging as model time horizons grow, potentially to hundreds of hours.

• Quality assurance for existing tasks. Once a task has been developed, you will verify that it's actually solvable as specified, and that the model is given (only) the information it needs.

• Baselining and scoring tasks. Where helpful, you may be asked to baseline tasks within your domain of expertise, and/or score task completions from AIs or human baseliners.

• Improving task development infrastructure. We're always improving our processes. Strong candidates will notice when existing workflows are inefficient or produce low-quality output, and take responsibility for improving them.

Skills we're looking for

Software engineering: You have several years of experience working on complex projects and codebases.

Evaluations: You have experience building hard (ideally agent-based) AI evaluations (e.g. RE-Bench, HCAST, SWE-bench Verified, Cybench, GPQA), ideally using the Inspect framework.

High attention to detail: You read closely, spot misspecifications and ambiguity, and pay attention to fiddly minutiae.

(Nice to have) Familiarity with METR infrastructure: Prior experience with Hawk, and familiarity with the methodology behind our Time Horizons work, is a plus.