Trillium Labs launches nonprofit for open AI safety research

Two AI researchers have launched Trillium Labs, a nonprofit dedicated to conducting high-stakes AI research transparently, publishing experimental details for outside scrutiny. The founders argue this open approach is crucial for mitigating risks posed by powerful frontier models. Nathan Lambert and Tom Zick will focus on areas like recursive self-improvement and reinforcement learning, which are often kept secret by major labs such as OpenAI and Anthropic. They believe closed development hinders scientific oversight and replication, while open methods could help the community understand and manage emerging AI behaviors. The lab has secured undisclosed funding from Schmidt Sciences and Halcyon Futures, aiming to raise $40-100 million total. It plans to spend $30 million on training over the next 18 months, hoping to add nuance to the debate over open versus restricted AI development.