A newly established research center will release open-source post-training recipes and shared infrastructure to investigate how the most powerful AI models behave. The organization intends to study reward-hacking and multi-agent systems through a transparent scientific process to ensure that frontier-model research does not become restricted to closed corporate environments.

Founded by Nathan Lambert and Tom Zick, Trillium Labs is based in the Bay Area and Cambridge, Massachusetts. The non-profit is currently recruiting full-time employees and student interns while seeking additional compute resources, having received initial funding from Halcyon Futures and Schmidt Sciences.