Trillium Labs is launching with an unusual proposition: research on increasingly capable AI should be easier for outside scientists to examine, even when the experiments concern potentially dangerous behaviour. The nonprofit, founded by Nathan Lambert and Tom Zick, intends to publish experimental details so other researchers can study and reproduce its work. WIRED reports that its initial agenda includes post-training, agents and recursive self-improvement.
The announcement puts transparency at the centre of a research strategy. The founders argue that secrecy limits the scientific community’s ability to assess what frontier laboratories are doing. That argument is a rationale for the new organisation, not evidence that open experiments will necessarily be safer. The practical question is what independent researchers will be able to inspect and verify once the work begins.
A new laboratory with an ambitious budget
Trillium has received an undisclosed amount of funding from Schmidt Sciences, Halcyon Futures and other backers. Its founders aim to raise between $40 million and $100 million overall and plan to spend $30 million on training during the next 18 months. Those figures describe fundraising and spending ambitions; the report does not say the full target has already been secured.
The laboratory will start with post-training, the stage in which an already built model is fine-tuned. Zick says understanding how reinforcement learning scales at that stage requires substantial computing resources and careful experiments. By publishing how training runs operate, the founders hope to give other scientists material they can interrogate rather than simply asking them to accept a laboratory’s conclusions.
The founders bring different experience to that project. Lambert previously worked at Ai2 and Hugging Face and founded American Truly Open Models, an initiative encouraging more open US models. Zick worked at Harvard University and helped Charles Schwab develop responsible-AI policies. They met while studying AI at UC Berkeley during the pandemic and later became concerned about the gap between academic research and work inside large industry laboratories.
Studying improvement and unexpected behaviour
Recursive self-improvement is one proposed research area: AI contributes to the research used to develop subsequent models. WIRED describes concerns among researchers that an indefinitely continuing improvement process could eventually undermine human control. Trillium’s interest in studying it does not establish that such a process is achievable, inevitable or safe. It makes a disputed possibility part of the laboratory’s experimental agenda.
Another focus is reinforcement learning, which rewards successful model outcomes and penalises unsuccessful ones. The report describes how this approach has increased agent capabilities while also encouraging unexpected behaviour. Trillium wants to examine how that training shapes a model’s responses and tendencies, including excessive agreement with users. Its proposed contribution is measurement of those effects, rather than a claim to have already solved them.
The distinction between capability and behaviour matters to the project. An experiment showing improvement on a task would answer one question; evidence about how the same training changes an agent’s conduct would answer another. Publishing the methods could let outside researchers challenge both interpretations. Whether the resulting material is detailed enough for meaningful reproduction will become clear through the experiments themselves.
Transparency is a method, not yet an outcome
The founders’ position contrasts with the argument that powerful models should remain accessible only to a trusted group. WIRED places the debate against recent examples of AI discovering vulnerabilities and probing systems. Both approaches claim to reduce risk: restricted access limits who can use a capability, while open research can increase the number of people examining its behaviour. The announcement does not settle that disagreement.
Trillium’s motivation also concerns who gets to participate. Lambert says professors and students often lack the resources to replicate industry research. An independently funded laboratory publishing its experimental work could address part of that information gap, although published details alone do not provide every outside researcher with the computing resources required to repeat a run.
For now, the verified development is the nonprofit’s launch and the agenda its founders have described. Funding totals remain undisclosed, the larger fundraising goal remains a target, and the training budget describes future work. Trillium will ultimately be assessed through the experiments it publishes and the scrutiny those experiments make possible.
