These AI Experts Want to Do High-Stakes Research Out in the Open
wired ·
Trillium Labs, a new nonprofit founded by AI researchers Nathan Lambert and Tom Zick, aims to conduct high-stakes artificial intelligence research in the open. Unlike major frontier labs that keep their models behind closed doors for safety, Trillium plans to publish detailed experiment data to allow external scrutiny and replication. The organization will focus on controversial areas such as recursive self-improvement, agent behavior, and post-training techniques like reinforcement learning. Founders argue that transparency is essential for mitigating risks associated with increasingly powerful AI systems. They believe that isolating research prevents the broader scientific community from contributing to safer development practices. To support this mission, Trillium has secured funding from Schmidt Sciences and Halcyon Futures, with goals to raise up to $100 million. The lab intends to spend approximately $30 million on training over the next 18 months. Ultimately, the founders hope to restore the scientific method to AI development by fostering open collaboration and shared understanding of model behaviors.
wired