A new nonprofit wants the most delicate frontier AI research conducted in public. Trillium Labs, launched by Nathan Lambert and Tom Zick, will publish the details of experiments into agents and recursive self-improvement so outsiders can examine and reproduce the work.
The two met as Berkeley graduate students, struck by how far industry research had drifted from academia. Professors and students, they noticed, could not reproduce what big labs were building for lack of resources. Lambert’s background runs through Ai2 and Hugging Face, and he started the American Truly Open Models push; Zick passed through Harvard and helped Charles Schwab write policies for responsible AI.
Their target list starts with post-training, the tuning that follows a model’s initial build. It extends to recursive self-improvement, in which AI contributes to designing its own successors, and to how reinforcement learning molds a model’s temperament, sycophancy included.
Momentum for the debate is real. An Anthropic researcher left this month warning that self-improvement could end human control, while openness has its own camp: Xiaomi published live details of a training run, and Stanford is pretraining a model called Marin in the open. Lambert says closed development reverses the scientific method.
