Lambert and Zick launch Trillium Labs to open up AI training research

Nathan Lambert and Tom Zick have launched Trillium Labs, a nonprofit that will publish AI training experiments for outside researchers to inspect and reproduce. Its first focus is post-training—the work done after a model’s initial training to refine its capabilities and behavior.

The founders’ argument is that researchers can increasingly observe what advanced models do without being able to investigate the training decisions behind that behavior. In their launch statement, they describe two obstacles: commercial incentives to keep methods private, and computing and engineering costs that prevent academic researchers from reproducing the work independently.

Opening the training recipe

Trillium says it will release complete post-training recipes: data, code, evaluations and intermediate checkpoints, or saved versions of models during training. Those materials are intended to let other researchers follow how behavior develops, test changes to the training process, investigate failures and adapt models to problems the original team did not anticipate.

The nonprofit structure is central to that promise. The founders say it will let them run controlled experiments, document failed training runs and publish findings that challenge their own assumptions, without needing to protect intellectual property. The commitment extends beyond releasing a finished model to exposing the work that produced it.

Capabilities and behavior, studied together

In interviews with WIRED, the founders outlined a research program that includes potentially risky areas of AI development. They want to examine not just whether training makes systems more capable, but how it changes their behavior. The planned work includes:

• Reinforcement learning: training that rewards good results and penalizes bad outcomes. Trillium will study how it affects capabilities and behavior, including excessive agreement with users.
• Recursive self-improvement: using AI to contribute research toward developing new models. The prospect of continuing improvement and loss of human control has alarmed some AI researchers.
• AI agents: another planned research area, alongside the lab’s investigation of how reinforcement learning shapes model character and unexpected behavior.

To understand something like how reinforcement learning scales in post-training, you need significant compute and a lot of careful experimentation

Tom Zick, speaking to WIRED

A training budget, not a completed fundraise

The founders told WIRED they aim to raise $40 million to $100 million in total and plan to spend $30 million on training over the next 18 months. Those figures describe a fundraising target and spending plan, not the amount already secured.

Trillium has received support from Halcyon Futures and Schmidt Sciences; the sum raised is undisclosed. Its launch statement says it is still fundraising, hiring and searching for computing resources. The founders want backers with different views about AI’s likely trajectory to support a shared scientific foundation.

The lab’s stated measure of success is how much independent research its resources enable. Its ambition is therefore not simply to publish its own findings, but to give other teams enough material to question, reproduce and extend them.