Trillium Labs has begun operations as a non-profit entity dedicated to advancing artificial intelligence research through open science principles. The organization intends to publish details of its experiments, particularly those involving potentially risky areas like AI self-improvement and the complex behaviors of AI models. This approach contrasts with the common practice in many frontier AI labs where sensitive research is kept private.
The founders, Nathan Lambert and Tom Zick, established Trillium Labs to counter what they describe as an increasingly closed-off nature of AI research. Lambert stated that keeping research contained within specific labs limits opportunities for broader scientific review and innovation from the wider community. The lab plans to initially concentrate on post-training improvements for large AI models and the process of recursive self-improvement (RSI). RSI, where AI models develop new models, has been a source of concern regarding human oversight. By publishing details on how reinforcement learning is applied in these processes, Trillium Labs aims to provide new insights into AI model evolution.
Trillium Labs is seeking funding and staff, with initial support from Schmidt Sciences and Halcyon Futures. The organization has plans to invest approximately $30 million in training over the next 18 months and aims to raise between $40 million and $100 million in total. In addition to research, the lab plans to develop research infrastructure to study topics such as reward hacking and multi-agent systems.
The organization's commitment to open science extends to making its research reproducible. This transparency aims to allow other scientists to verify findings and build upon the work, fostering a more collaborative and scrutinized development environment for AI. The initiative is seen by some as a move towards actual open science in the field. A related paper co-authored by Nathan Lambert, concerning Context Language Models (CLMs), was submitted to arXiv on September 30, 2026, involving collaborators from the University of Washington, Meta Superintelligence Labs, and MIT.