Two AI industry scientists have launched a nonprofit to research potentially dangerous artificial intelligence technologies in full public view, according to a report published today by WIRED. Nathan Lambert and Tom Zick founded Trillium Labs to work on controversial areas like recursive self-improvement and AI agents while publishing detailed experimental results for outside scientists to examine and reproduce. The pair believe transparency, not secrecy, is the best path to understanding and controlling advanced AI systems.
The nonprofit, which launched today, will initially concentrate on post-training, the process of fine-tuning large models after they've been constructed. Another central focus will be recursive self-improvement, a technique where AI systems contribute to developing new models—a prospect that has alarmed many researchers because ongoing progress could theoretically continue without end, potentially leading to a loss of human control. The organization will also examine how reinforcement learning, which rewards models for good results and punishes poor outcomes, shapes AI capabilities and behavior. Trillium Labs has secured funding from Schmidt Sciences, Halcyon Futures, and other backers in an undisclosed amount, with founders targeting $40 million to $100 million in total fundraising and planning to spend $30 million on training over the coming 18 months.
Lambert, who previously worked at AI research lab Ai2 and founded the American Truly Open Models initiative, told WIRED that frontier AI labs' secrecy "reduces the community's ability to scrutinize ideas and contribute new approaches." Zick explained that publishing details of reinforcement training runs could generate unexpected insights as external researchers examine the work. According to the report, Lambert believes "the scientific method and careful measurement of recent events is the best way to understand new behaviors of AI models." The founders say they aim to inject much-needed nuance into debates about AI development strategy.
The AI industry is currently divided over whether powerful models should be kept inside company labs with limited access or released openly for broader scrutiny. Companies like OpenAI and Anthropic restrict their most powerful models to app or API access, limiting transparency about how systems are built and how they behave. The report notes that Chinese companies and some academic groups have taken the opposite approach—the Chinese company Xiaomi recently published live details of a major training run, while Stanford researchers are pretraining the AI model Marin in the open. Those favoring limited access argue that frontier models can now automate the discovery of software vulnerabilities and hack into systems, so keeping that power restricted to trusted parties is essential. Lambert and Zick's camp counters that shared understanding of risks makes everyone safer.
The founders hope Trillium Labs will demonstrate an alternative path for AI development at a moment when the field has grown disconnected from academic research. Lambert says professors and students often can't replicate work happening inside major company labs because they lack necessary resources. The issue of recursive self-improvement gained mainstream attention earlier this month when an Anthropic researcher departed the company and warned the technique could pose an existential threat to humanity. By making their experiments transparent, the nonprofit aims to show that open scientific inquiry, not corporate secrecy, offers the best route to building AI systems safely. The transparency model requires patience and tolerance for outside criticism, but it may prove essential as AI capabilities continue to advance. If influential labs remain committed to opacity, the industry risks repeating past mistakes where lack of scrutiny allowed problems to compound undetected.