In a world where closed labs are valued in the tens and hundreds of billions of dollars, two researchers decide to do the opposite: publish data, code, and intermediate post-training checkpoints. Founders Nathan Lambert and Tom Zick are launching the nonprofit Trillium Labs to restore to AI research the scientific openness that once created the entire industry.
Modern AI grew out of a shared scientific commons — papers, benchmarks, and code that universities and companies shared. But as commercial value grew, post-training and reinforcement learning became closed: publishing full recipes is now expensive, since they directly affect product and IPO valuations. Anthropic plans to raise up to 100 billion dollars, and every secret method for a reliable agent is a competitive advantage.
Trillium Labs intends to spend up to 30 million dollars on training over the next 18 months, with an overall fundraising goal of 40–100 million. The money has already come from Schmidt Sciences and Halcyon Futures. Unlike commercial labs, a nonprofit structure makes it possible to document failed runs, publish every stage of experiments, and not protect IP. This gives independent researchers the ability to test hypotheses that corporations would never publish.
The focus is on post-training, reinforcement learning, agents, and recursive self-improvement. This is precisely where model behavior takes shape most unpredictably, and where closedness hinders the collective search for solutions. Lambert and Zick, who previously worked at the Allen Institute for AI and Hugging Face, see this as a return to the model that once allowed the community to accelerate progress many times over.
The economic paradox is obvious: the more valuable the knowledge, the stronger the temptation to hide it, yet it is precisely openness that has historically created the greatest value for the entire industry. Trillium Labs is betting that the long-term payoff from collective intelligence will surpass the short-term valuations of closed giants.
In the end, the question is not only about technology, but also about how we distribute access to the most powerful tool of the era. Open recipes could become those "seeds" that give growth to the whole forest, and not just to a few tall trees.
