AI researchers Nathan Lambert and Tom Zick launched Trillium Labs, a non-profit research organization dedicated to public, open-science AI development. Announced on October 2, 2026, the lab plans to publish full experimental details, training methods, and post-training data. The organization will focus on high-stakes, safety-critical domains such as recursive self-improvement (RSI) and autonomous agent behaviors.

The venture directly challenges the closed development practices favored by leading labs like OpenAI and Anthropic, which restrict frontier model access to closed APIs. Trillium’s founders argue that open research and peer replication are required to identify safety vulnerabilities and prevent unintended risk as models gain autonomous capabilities.

Trillium Labs will initially target post-training fine-tuning and transparent methodology to bridge the growing resource gap between commercial labs and academic researchers, joining open efforts like Stanford’s Marin project and Ai2.

Why it matters

  • Access open post-training research and methodologies to audit and fine-tune proprietary autonomous agent pipelines.

  • Monitor open safety frameworks around recursive self-improvement to prepare for emerging agent risk regulations.

  • Leverage transparent academic research as an alternative to proprietary API lock-in from major frontier labs.

Source: wired.com