Trillium Labs opens nonprofit AI lab to publish the training methods frontier models keep closed

The new research organization will release open post-training recipes, code, data, evaluations and model checkpoints, with initial support from Halcyon Futures and Schmidt Sciences

Trillium Labs has been established as a nonprofit research organization focused on making frontier AI post-training methods openly available

Trillium Labs has been established as a nonprofit AI research lab focused on opening up the methods used to train frontier AI systems, beginning with the increasingly important post-training stage.

Co-founded by AI researcher Nathan Lambert and Tom Zick, the organization plans to publish complete post-training recipes rather than only releasing finished models or research results. That includes the underlying data, code, evaluations and intermediate checkpoints, allowing outside researchers to examine how model behavior changes during training and build on the work independently.

Trillium argues that access to this level of detail has become increasingly limited as the commercial value of frontier AI research has grown. It points to earlier periods of AI development when researchers inside companies including Google and OpenAI published work on technologies such as the transformer architecture and reinforcement learning from human feedback with enough methodological detail for researchers elsewhere to investigate and extend it.

The organization says that has become harder as frontier AI development has moved toward expensive post-training programs, reasoning systems and AI agents.

“We believe these problems require a scientific community with the resources to investigate them independently,” Trillium states. “A handful of closed research programs cannot provide the diversity of questions, methods, and perspectives that a technology this consequential needs.”

Opening up AI post-training

Post-training is the stage after a model’s initial training in which developers can further shape its behavior, capabilities and responses. Trillium plans to make its work in this area fully open so researchers can test interventions, investigate failures and adapt models for other research questions.

The lab says its nonprofit structure will allow it to publish failed experiments as well as successful ones, run controlled studies and release findings without needing to protect the intellectual property generated by the research.

Its work will build on open AI research from organizations including the Allen Institute for AI, EleutherAI, OpenAthena, Nvidia and Hugging Face.

Lambert wrote on LinkedIn that Trillium intends to expand beyond post-training into open infrastructure for researching areas including recursive self-improvement, reward hacking and multi-agent systems.

“We're built around the theory of change that you need more eyes to solve hard technical problems,” he wrote. “We have faith in the scientific methods and communities that humanity has built, and worry that AI is becoming too closed to utilize them.”

Funding, research and recruitment

Trillium has begun with support from Halcyon Futures and Schmidt Sciences and is seeking further funding from a wider group of backers.

The organization says it wants supporters with different views on AI development and risk to contribute to shared research infrastructure that can be used to test competing claims about frontier systems.

Lambert said Trillium is also looking for additional computing resources and recruiting for full-time positions, student collaborations and internships. Its offices are based in the San Francisco Bay Area and Cambridge, Massachusetts, with remote work also available.

The lab has named Thomas Wolf, Hanna Hajishirzi, Graham Neubig and Bryan Catanzaro as advisors.

Its planned model releases will form one part of the work. Trillium says the longer-term objective is to create openly available training methods and research infrastructure that other scientists can repeatedly use to investigate frontier AI.