Frontier artificial intelligence labs routinely guard their inner workings and training runs behind tightly sealed corporate walls. Challenging that entrenched secrecy, two prominent AI researchers have established a new nonprofit organization designed to execute critical machine learning research in public view. Founded by Nathan Lambert and Tom Zick, Trillium Labs intends to probe complex and controversial frontiers of AI, such as recursive self-improvement (RSI) and autonomous agents, with radical transparency. The core operational model requires publishing granular experimental logs and technical methodology so that academic and independent scientists can evaluate and reproduce the findings.
Lambert argues that the extreme opacity preferred by leading frontier builders actively stifles the wider scientific ecosystem from interrogating hypotheses and contributing fresh approaches. Allowing independent specialists to observe precisely how advanced models are assembled and aligned represents an indispensable prerequisite for containing long-term safety risks.
“The current closed trajectory of frontier AI development is taking us a step backwards.”
At present, state-of-the-art models developed by organizations like OpenAI and Anthropic are strictly gated behind end-user interfaces or proprietary application programming interfaces (APIs). This structural barrier comes at a significant cost to public transparency regarding internal model architecture, training data hygiene, and downstream operational quirks.
The Clash Between Open Distribution and Walled Gardens
In contrast to western proprietary labs, alternative institutions, notably across China, continue to distribute powerful open-weight models that practitioners can host and run on their own computing infrastructure. Chinese enterprise Xiaomi recently showcased this philosophy by publishing live telemetry and operational details from a major model training run. Concurrently, academic researchers at Stanford University are conducting the pretraining phase for the Marin model entirely in the open.
This divide has ignited intense debate over the safest path forward, primarily because contemporary frontier models possess potent, dual-use capabilities. Advanced systems can autonomously discover novel software vulnerabilities, probe defense perimeters, and execute exploit pathways. Heightened scrutiny has intensified in the wake of high-profile hacking incidents worldwide.
Advocates of closed, gated access argue that concentrating such powerful capabilities within the custody of a strictly vetted few remains vital for global security. Conversely, researchers aligned with Lambert and Zick maintain that shared awareness of vulnerabilities and collaborative risk reduction deliver vastly superior collective protection.
Founding Origins Across UC Berkeley and Corporate Labs
Nathan Lambert brings a long track record in open-source AI development. His background includes tenure at Ai2, an institution renowned for releasing extensive model training data and methodology. Lambert also contributed to Hugging Face, authored an influential technical publication, and spearheaded the American Truly Open Models initiative to rally American tech firms behind genuinely open architectures. Tom Zick previously worked at Harvard University, where his research informed governance structures, and assisted Charles Schwab in formulating responsible AI operational guidelines.
The co-founders originally connected via Zoom during the COVID-19 pandemic while both were graduate students at UC Berkeley investigating artificial intelligence. The vision for Trillium Labs emerged directly from observing a widening chasm between industry developments and academic computer science departments. University faculty and graduate students frequently encounter insurmountable computing resource bottlenecks that prevent them from reproducing breakthroughs engineered inside deep-pocketed corporate laboratories.
Priorities in Post-Training, Reinforcement Learning, and RSI
According to Zick, Trillium Labs will dedicate its opening research cycles to post-training workflows, specifically the refinement and alignment of expansive foundation models after initial pretraining is complete. A second major focus centers on recursive self-improvement, the paradigm where an AI system actively conducts research to architect subsequent model generations. The potential for such autonomous compounding to escape human steering mechanisms has stoked intense concern across the engineering field, gaining wider prominence earlier this month when an Anthropic researcher resigned and warned that unchecked RSI could trigger existential dangers for humanity.
The organization will also investigate how reinforcement learning, the protocol that rewards desired outputs and penalizes unwanted behaviors, alters model capability. While reinforcement learning has turbocharged agentic capabilities, it has also produced erratic and emergent behaviors. The research team intends to chart how reinforcement training shapes model personalities and behaviors, particularly when systems develop problematic traits like deceptive sycophancy.
Funding Horizons and the Scientific Path Ahead
Zick emphasized that untangling the scaling dynamics of reinforcement learning during post-training demands substantial compute infrastructure and methodical, disciplined testing. Releasing the underlying data of these reinforcement cycles could expose vital structural insights that remain concealed within private corporate servers.
Trillium Labs has secured preliminary capital from Schmidt Sciences, Halcyon Futures, and several additional philanthropic entities. The co-founders are aiming to secure between $40 million and $100 million in total funding, with immediate blueprints allocating $30 million toward computational training over the upcoming 18 months.
Tim Fist, who serves as director of emerging technology policy at the Institute for Progress think tank, voiced strong backing for the effort, emphasizing his commitment to greater transparency across machine learning research and development. Lambert and Zick believe that introducing transparent research will elevate the broader public debate around frontier AI architecture, shifting the conversation away from isolated corporate viewpoints toward systematic, peer-reviewed measurement.
