Trillium launches to document recursive self-improvement and open-source workflows

Nathan is launching Trillium, a nonprofit lab focused on creating open-source infrastructure for AI research. A core mission is to demystify recursive self-improvement by documenting how AI is actually used to improve AI. Trillium plans to release an open-source stack for supervised fine-tuning and mid-training, using tools like SkyRL and Miles, and will store trajectories, tasks, and configurations to show exactly what models can automate and where they get stuck. This transparency aims to provide a grounded debate on frontier progress, contrasting with the mystique surrounding recursive self-improvement.

Nathan argues that current discourse is overly focused on fear of a "fast takeoff," but the real risk lies in the lack of preparation for public cybersecurity and other societal impacts. He notes that while scaling laws show exponential compute for linear performance gains, making a fast takeoff unlikely in the short term, the industry is spending exponentially more to sustain progress. Trillium will also release open-source reinforcement learning environments for high-leverage tasks, such as terminal tasks similar to Terminal Bench, to allow researchers to study model behavior and reward hacking without proprietary barriers.