Trillium Labs pushes high-stakes AI research into the open
Trillium Labs, a nonprofit co-founded by Nathan Lambert and Tom Zick, aims to move high-stakes AI research — on topics such as self-improvement, agents and model behavior — out of closed frontier labs by sharing experimental details so outside researchers can inspect and reproduce results. Lambert argues that leading systems, especially models from OpenAI and Anthropic, are typically available only through apps or APIs, leaving outsiders unable to understand how they are built or how they actually behave, and that long-term enclosure by a few labs weakens the community's ability to evaluate new ideas and propose alternatives. On the same day, Geoffrey Hinton and other leading AI researchers published a paper warning of a possible 'intelligence explosion' and calling on policymakers to require standardized disclosure by AI companies; Trillium Labs researchers had also worked with Meta, the University of Washington and MIT on a Context Language Models architecture.
Conducting high-stakes research in the open could let outside experts scrutinize model design and fine-tuning earlier, helping the community assess risks around issues like self-improvement.