Lambert, Zick Launch Trillium Labs, Targeting $40 Million to Open AI Risk Research
Updated
Updated · WIRED · Oct 2
Lambert, Zick Launch Trillium Labs, Targeting $40 Million to Open AI Risk Research
1 articles · Updated · WIRED · Oct 2
Summary
Trillium Labs launched as a nonprofit to publish replicable AI experiments, with founders Nathan Lambert and Tom Zick arguing that openness can improve scrutiny of frontier-model risks.
The lab will start with post-training work and research on recursive self-improvement and reinforcement learning—areas tied to more capable agents, unexpected behavior and fears of losing human control.
Lambert and Zick say closed corporate labs have widened the gap with academia, leaving professors and students unable to reproduce key results because they lack access and compute.
Backed by Schmidt Sciences, Halcyon Futures and others, the group has raised an undisclosed amount, aims for $40 million to $100 million overall, and plans to spend $30 million on training in 18 months.
The launch lands in a broader fight over AI access: companies such as OpenAI and Anthropic restrict frontier models, while some Chinese firms and Stanford's Marin project have pushed more open development.