Some major AI companies believe that restricting access to AI models can prevent potential chaos and misuse. Nathan Lambert and Tom Zick, two scientists in the field, argue for a different approach. They founded a nonprofit, Trillium Labs, to conduct AI research transparently, including in areas like recursive self-improvement (RSI) and agent-based systems. Their goal is to publish experimental details to allow external scientists to replicate and scrutinize their work.
Lambert states that the secrecy surrounding frontier AI labs limits the community's ability to evaluate ideas and contribute new solutions. He emphasizes that transparency in model development is essential for risk mitigation. "Over the past few millennia, humanity has had the scientific method in our toolbox as a way to mitigate harms and build better futures," Lambert said. He believes the current trend towards secrecy in AI development is regressive.
Leading AI models from companies like OpenAI and Anthropic are typically accessed through applications or APIs, which often lack transparency regarding their construction and behavior. In contrast, some companies, particularly in China, provide models that can be downloaded and run on personal hardware. For instance, Xiaomi recently shared details of a significant training run for one of its models, and Stanford researchers are pretraining the AI model Marin openly.
The industry is currently divided over the best approach to managing powerful AI models, which can automate the discovery of software vulnerabilities and conduct system probing. Advocates of limited access argue that it is crucial to keep such capabilities within a trusted circle, while Lambert and Zick contend that a shared understanding of risks benefits everyone.
Lambert has previously worked at Ai2, a research lab known for its open approach to AI, and has been involved in initiatives encouraging the release of open models. Zick, who has worked at Harvard University, has contributed to policies around responsible AI.
The two met during the COVID-19 pandemic while studying at UC Berkeley and recognized the disconnect between industry AI research and academic work, which often hinders replication due to resource constraints.
Trillium Labs, which officially launched on October 2, 2026, will initially focus on post-training of large models and exploring RSI, a method that allows AI to contribute to its own research. This concept has raised concerns among AI researchers about the potential for losing human control over ongoing advancements. The issue gained attention when an Anthropic researcher warned that RSI could pose an existential threat.
The nonprofit will also investigate reinforcement learning, a technique that rewards models for positive outcomes and penalizes them for negative ones. Zick notes that understanding how reinforcement learning operates in post-training requires substantial computational resources and careful experimentation. She believes that publishing details of reinforcement training could yield valuable insights as external researchers analyze the findings.
Trillium Labs has secured undisclosed funding from Schmidt Sciences, Halcyon Futures, and others, with plans to raise between $40 million and $100 million in total and allocate $30 million for training over the next 18 months. Tim Fist, director of emerging technology policy at the Institute for Progress, expressed support for increased transparency in research and development.
Ultimately, Lambert and Zick aim for Trillium Labs to add complexity to the ongoing discourse on AI development. "We’re in an era of AI discourse dominated by a few world views," Lambert said. "We believe that the scientific method and careful measurement of recent events is the best way to understand new behaviors of AI models."