A nonprofit startup, Trillium Labs, plans to push frontier AI research into the open, publishing experiment details to invite external study and replication. Founders Nathan Lambert and Tom Zick argue that making work public could strengthen the scientific method in AI development and help mitigate potential harms associated with powerful models. The initiative focuses initially on -training work—fine-tuning large models after their initial build. It will also study recursive self-improvement and the use of reinforcement learning to shape model behavior, including how such training regimes affect a model's reliability and safety. Zick notes that understanding how reinforcement learning scales in -training requires significant compute and careful experimentation, and that sharing process details could yield valuable insights for researchers outside big labs. Lambert emphasizes that current closed trajectories in frontier AI development limit communal scrutiny and collaboration, potentially hindering progress toward safer outcomes. Trillium Labs has raised funds from Schmidt Sciences, Halcyon Futures, and others, with a target to raise a total of 40 to 100 million dollars and plans to spend about 30 million on training over the next 18 months. The founders say they hope the open approach will help integrate academic and industry perspectives, a contrast to how some leading models remain accessible only via apps or APIs. Lambert previously worked at Ai2 and Hugging Face and has advocated for releasing more open models through initiatives like American Truly Open Models. Zick has experience at Harvard and helped shape responsible AI policies at Charles Schwab. The two met as UC Berkeley graduate students and were inspired by the perceived disconnect between industry experiments and academic replication. Their effort signals a broader debate over how to balance openness with safety as models become increasingly capable, including the prospect that ongoing progress could outpace human control. Proponents contend transparency supports accountability; skeptics warn of risks from revealing too much about sensitive research pathways. Trillium Labs plans to evaluate how -training and reinforcement learning affect model personality and behavior, aiming to shed light on when a model becomes overly compliant or unpredictable. The initiative promises to contribute nuance to AI policy discussions, insisting that measured, openly shared experiments can illuminate better paths forward for the field.