New Nonprofit Pushes for Transparent, Replicable AI Research on Risky Frontiers

Nathan Lambert and Tom Zick argue that today's leading AI companies have it backwards. Rather than locking powerful models away from the public to minimize risk, the pair founded Trillium Labs, a nonprofit that will tackle sensitive research areas—including recursive self-improvement and autonomous agents—by openly sharing experimental methodology and results.

The founders believe that secrecy in frontier AI research prevents the broader scientific community from catching mistakes, validating findings, or offering alternative approaches. "Over the past few millennia, humanity has had the scientific method in our toolbox as a way to mitigate harms and build better futures," Lambert tells WIRED, contrasting this with what he sees as the closed nature of today's AI development.

The Open vs. Closed Divide#

Major labs at companies like OpenAI and Anthropic restrict access to their most capable models through APIs and applications, keeping details of training and behavior under wraps. Other players—particularly Chinese firms and academic efforts like Stanford's Marin project—have chosen a more permissive path, releasing models that can run on local hardware or publishing training details live.

The stakes feel higher now. Modern AI systems can discover software vulnerabilities automatically and penetrate computer networks, raising alarms about whether such power should be held by a narrow group or distributed more widely.

Focus Areas and Early Support#

Trillium Labs plans to concentrate initially on post-training refinement and recursive self-improvement—the prospect that AI systems might autonomously accelerate their own development, potentially outpacing human oversight. The lab will also examine how reinforcement learning (rewarding good outputs, penalizing bad ones) shapes AI capabilities and behavior, including unwanted traits like excessive agreeableness or unexpected actions.

The organization has already secured funding from Schmidt Sciences, Halcyon Futures, and others, with founders targeting $40–100 million overall and earmarking $30 million for computational training over the next 18 months.

Lambert's Track Record#

Lambert previously led research at Ai2, a lab known for publishing model training details and data openly. He has also worked at , maintained a , and launched the initiative to push U.S. companies toward greater transparency. Zick brings expertise from Harvard and experience developing AI governance frameworks for Charles Schwab. The two met as UC Berkeley graduate students during the pandemic and recognized that academic researchers often cannot reproduce industry work due to resource constraints.