Fewshot Corp.

We are working in reward hacking detection and mitigation. We are interested in building a product that helps researchers and developers implement their own mitigation. We have a concrete research agenda to study reward hacking tasks to complete our empirical study measuring how reward visibility affects hacking behavior, demonstrate whether RL training systematically amplifies reward hacking, and establish actionable guidelines for test design. We believe a commercially viable solution is possible and plan to build a company dedicated to AI training safety, with the ultimate goal to institutionalize safety and build an institution dedicated to independent safety evaluation.

Our Work

Fewshell

Mobile assistant for DevOps, On-Calls, and AI Researchers. Safely manage your infrastructure from anywhere.

In The Wild