Trillium Labs Launches to Open Up High-Stakes AI Research

AI researchers Nathan Lambert and Tom Zick have launched Trillium Labs, a new nonprofit research organization dedicated to conducting high-stakes artificial intelligence experiments in the public eye. Backed by funding from Schmidt Sciences, Halcyon Futures, and other donors, the group aims to raise between $40 million and $100 million in total, with plans to spend $30 million on model training over the next 18 months. The initiative represents a direct challenge to the closed-door development strategies favored by industry leaders like OpenAI and Anthropic.

Trillium Labs will focus its initial efforts on post-training methodologies, specifically fine-tuning large models after their initial creation. The founders also plan to investigate recursive self-improvement (RSI)—a controversial process where AI models contribute to their own development—and the ways reinforcement learning affects the behavior and character of AI agents. Lambert, who previously worked at Ai2, Hugging Face, and founded the American Truly Open Models initiative, teamed up with Zick, a former Harvard researcher who helped Charles Schwab develop responsible AI policies, to bridge the growing resource gap between corporate labs and academic researchers.

This open approach contrasts sharply with the current industry trend of restricting access to protect against risks like automated hacking. However, some organizations are already pushing for transparency; Stanford researchers are currently pretraining an AI model called Marin in the open, while Chinese tech firm Xiaomi recently shared live details of a major training run. For AI practitioners and academic researchers, Trillium Labs provides a rare opportunity to study the mechanics of advanced post-training and reinforcement learning without needing the massive proprietary budgets of commercial tech giants. Lambert argues that the current closed trajectory of frontier AI development is taking the scientific community "a step backwards" in its ability to mitigate risks.