Trillium Labs reportedly launched on October 2, 2026, as a nonprofit founded by Nathan Lambert and Tom Zick. The founders intend to publish experiment details so outside researchers can scrutinize and replicate the work, putting openness at the center of the lab’s stated approach.
What Trillium Labs plans to study
The lab’s initial focus is post-training: additional work on a model after it has been built, including fine-tuning. Its planned research also covers how reinforcement learning affects models’ capabilities and behavior, including unexpected behavior and sycophancy—when a model agrees too readily with a user.
Other intended topics include AI agents and recursive self-improvement, the study of whether AI systems could help improve later versions of themselves. These are areas on the research agenda, not reported findings.
The debate over open and restricted AI research
Publishing experiment details could let more researchers examine results and try to replicate them, broadening scrutiny of how AI systems behave. That is the case for openness associated with Trillium Labs’ approach.
The counterargument is that limiting access to powerful models can keep potentially harmful capabilities in the hands of trusted users and reduce real-world risks while research continues. The debate turns on how much access and experimental detail to make public as AI capabilities are studied.
The founders’ financial plans
The founders aim to raise $40 million to $100 million in total. Separately, they plan to spend $30 million on training over 18 months. Trillium Labs has reportedly received funding from Schmidt Sciences, Halcyon Futures and other supporters.
