About Fleet Fleet studies how environments produce safe intelligence. We believe safety and economic value increasingly reinforce each other: as agents take on more consequential work, their value depends on whether they can be trusted with that responsibility.
About the role
About Fleet
Fleet studies how environments produce safe intelligence. We believe safety and economic value increasingly reinforce each other: as agents take on more consequential work, their value depends on whether they can be trusted with that responsibility.
As reinforcement learning takes a larger role in post-training, the design of tasks, environments, and feedback becomes increasingly consequential for downstream alignment. RL can reinforce sound judgment, but it can also amplify unintended behaviors. We believe studying and scaling the data and training signals that produce more aligned agents is one of the most impactful ways to contribute to AI safety.
Backed by Sequoia Capital, Menlo Ventures, BCV, and SV Angel.
About the Role
We’re looking for an Applied Research Scientist to work on robust reward design, implement comprehensive telemetry, and design production systems to help ensure Fleet's RL tasks make capable agents more aligned. You’ll work with data and model trajectories to help develop oversight systems to detect misaligned behaviors. You’ll also work directly with safety researchers at frontier labs to help co-design new RL environments and tasks for downstream alignment training and evaluation.
What You’ll Work On
Reward Design and Telemetry at Production Scale. Study how task and verifier design affect model-alignment incentives, then improve both.
Off-target behaviors from task assumptions and incentives. Investigate how unstated behavioral constraints, unsolvable tasks, and rewards that ignore how success is achieved affect how models behave during evaluation, across many different domains.
Telemetry and Verifier Generation Pipelines. Use the insights learned from the above to build 1) the telemetry needed for customers to have complete observability over undesirable model behaviors and 2) the scalable pipelines needed for producing verifiers that translate recorded measurements into qualitative behavioral categories. …
Official source: fleetai.com. Details can change; the platform's page is the final word. How we check listings.
