A new nonprofit called Trillium Labs is taking a deliberately open approach to some of artificial intelligence's most sensitive research areas. WIRED reports that industry scientists Nathan Lambert and Tom Zick founded the organization to investigate model post-training, AI agents and recursive self-improvement while publishing experimental details for outside researchers to study and reproduce.
The project enters a debate over whether increasingly capable AI should be studied behind closed doors or exposed to broader scientific scrutiny. According to WIRED, leading systems from companies such as OpenAI and Anthropic are generally available through applications or APIs rather than as downloadable models, limiting what independent researchers can inspect about their construction and behavior. In contrast, some organizations release models or development details more openly; WIRED points to a Xiaomi training run shared live and Stanford researchers' open pretraining work on a model called Marin.
Lambert and Zick argue that secrecy prevents academics and other experts from testing claims, repeating experiments and contributing alternative ideas. WIRED reports that the founders met remotely while both were graduate students at the University of California, Berkeley, during the Covid-19 pandemic. Their idea for Trillium grew from concern that academic researchers often lack both the access and computing resources needed to replicate work performed inside large corporate labs.
The founders bring experience from open-model research and responsible-AI policy. WIRED says Lambert previously worked at the Allen Institute for AI and Hugging Face, founded the American Truly Open Models initiative and writes a technical blog. Zick previously worked at Harvard University and helped Charles Schwab develop responsible-AI policies.
Trillium's initial technical focus will be post-training, the stage in which an already-built model is further refined. The nonprofit also plans to examine reinforcement learning, which uses rewards and penalties to shape model behavior. WIRED reports that the founders want to study how those techniques affect both capability and character, including unexpected behavior or excessive agreement with a user.
A more controversial research area is recursive self-improvement, described by WIRED as using AI to help conduct research that contributes to developing new models. Some researchers worry that repeated gains could become difficult for humans to control. Trillium's decision to investigate the subject openly therefore cuts against the view that access to potentially risky capabilities should remain limited to a small number of trusted organizations.
The founders' argument is that openness could improve risk management by allowing more experts to examine methods and results. Their planned publications are intended to let outside scientists scrutinize experiments and attempt replication. That approach does not establish that open publication will be safer in every case, and WIRED's report does not specify how Trillium will decide whether any result is too hazardous to release.
The undertaking will require substantial computing resources. WIRED reports that Trillium has received an undisclosed initial amount from Schmidt Sciences, Halcyon Futures and other backers. The founders aim to raise between $40 million and $100 million and say they plan to spend $30 million on model training during the next 18 months. Those figures describe goals rather than completed fundraising or spending.
Tim Fist, director of emerging technology policy at the Institute for Progress, told WIRED that he supports greater transparency in AI research and development. The report presents Trillium as an effort to bring more scientific measurement into a field whose most advanced work is increasingly concentrated inside well-funded companies.
For now, Trillium's significance lies more in its research model than in proven results. The organization launched on October 2, 2026, and WIRED does not report completed experiments, released models or demonstrated safety outcomes. Its test will be whether publishing detailed, compute-intensive work can meaningfully expand independent understanding without amplifying the risks that closed-lab advocates cite as the reason for restricting access.