Researchers Nathan Lambert and Tom Zick have launched Trillium Labs, a nonprofit that plans to publish experiments in potentially risky areas of artificial intelligence rather than keep the work behind closed doors. Its initial focus includes model post-training, autonomous agents and recursive self-improvement, in which AI contributes to developing later systems.
The founders argue that secrecy at frontier labs prevents outside scientists from testing claims, finding weaknesses and proposing safeguards. Trillium intends to disclose enough experimental detail for independent researchers to study and reproduce its work. Lambert previously worked at Ai2 and Hugging Face and founded an initiative promoting genuinely open US models. Zick has worked on responsible-AI policy at Harvard and Charles Schwab.
The project enters a live dispute over whether dangerous capabilities are safer when access is restricted. Closed-model developers point to abilities such as vulnerability discovery and automated intrusion as reasons for limiting release. Advocates of open research contend that wider scrutiny improves understanding and risk mitigation.
Trillium’s launch establishes an approach, not evidence that open publication will make advanced systems safer. The nonprofit will still need to define what it will release, how it will handle findings that could enable abuse and whether outside teams have enough computing resources to replicate its experiments.