Trillium Labs launches to pursue open frontier-AI science across post-training, RSI and multi-agent research

What changed

Nathan Lambert and Tom Zick introduced the nonprofit Trillium Labs on October 1 to rebuild a more open scientific commons around frontier AI. The lab will start with post-training and plans to publish recipes, failed runs and evaluation details, later expanding into recursive self-improvement, reward hacking and multi-agent systems. It names initial support from Halcyon Futures and Schmidt Sciences while continuing to raise funding and seek compute.