Nathan Lambert and Tom Zick launch Trillium Labs, a non-profit lab for open science in AI

Trillium Labs is the name of the new nonprofit lab announced on October 1, 2026 by Nathan Lambert, a researcher well known in the open AI ecosystem, and Tom Zick, whom he describes as "a longtime friend and collaborator." The project is presented with a specific goal: to foster open science around frontier AI, starting with post-training (the tuning phase after pretraining, which includes techniques such as reinforcement learning and instruction fine-tuning) and with the intention of later expanding into open infrastructure for studying phenomena such as recursive self-improvement (RSI), "reward hacking" (when a model exploits flaws in the reward function instead of solving the intended task) and multi-agent systems.

The announcement opens with a historical diagnosis: according to Lambert and Zick, modern AI is the child of a "scientific commons" built over decades across universities and industry, where knowledge circulated through papers, code and public benchmarks. They cite two specific examples that shaped the latest wave of generative AI: the paper "Attention is All You Need," which introduced the transformer architecture and came out of Google, and OpenAI's InstructGPT paper, which spelled out at the level of algorithmic design how to apply reinforcement learning from human feedback (RLHF). Both works, they note, allowed researchers outside those companies to examine, adapt and extend those techniques.

The text's central argument is that this tap of open research has been shutting off precisely when it matters most: as post-training has become decisive for reasoning and autonomous-agent capabilities, access to complete training recipes has grown scarce. The authors attribute this closure to simple economics: the cost of transparency has risen in step with the commercial value of that knowledge. As an illustration of the sums at stake, the article notes that, according to the Wall Street Journal, Anthropic's proposed IPO could raise up to $100 billion, a figure that would exceed the combined total of the Saudi Aramco, Alibaba and SoftBank Corp offerings (some $78 billion). A method that makes an agent more reliable, they reason, also makes it a better product, so publishing it amounts to handing competitors part of the knowledge underpinning those valuations.

On top of this commercial closure, the text says, comes a practical barrier: reproducing a serious training program demands so much compute and engineering that academic researchers cannot fill in the missing details on their own. The result is a combination they describe as dangerous: companies that may be entering one of the greatest periods of corporate growth in history, while developing a foundational technology whose effects reach far beyond their customers and shareholders, even as the scientific community loses its capacity for independent scrutiny of increasingly consequential training decisions.

Against that diagnosis, Trillium Labs makes a bet: that these problems — aligning the systems, securing them, spreading their benefits across society — require a scientific community with the resources to investigate them independently, because a handful of closed research programs cannot supply the diversity of questions, methods and perspectives a technology of this magnitude demands. Their concrete proposal is to build fully open post-training recipes: not just the resulting models, but the data, code, evaluations and intermediate checkpoints that let other researchers study how a model's behavior develops step by step.

The article explicitly places Trillium Labs as a continuation of work already being done by other open AI organizations — it mentions the Allen Institute for AI, EleutherAI, OpenAthena, Nvidia and Hugging Face — while arguing that this collective effort alone is not enough. Their differentiating argument is the nonprofit structure: with no intellectual property to protect, they say they can afford to invest in controlled experiments, document failed training runs, study how an intervention at one stage affects model behavior at later stages, and publish findings that even contradict their own starting hypotheses. They define their measure of success not in product terms but in how much independent third-party research their resources enable.

On backing and resources, the announcement says the project has launched with support from Halcyon Futures and Schmidt Sciences, and that they are looking to broaden funding toward a diverse coalition of funders who may disagree about which AI risks matter most or which scenarios are more likely, but who agree on sustaining a common scientific base. The text does not detail amounts for that initial funding or an overall budget for the lab. Lambert and Zick also announce that they are hiring (full-time roles and collaborations or internships for students), that they are still seeking funding and that they need access to compute, and they invite anyone interested to get in touch; Lambert notes he will be in San Francisco until October 9 for the launch.

The lab's name refers to trilliums, ephemeral spring flowers that bloom before the forest canopy closes: their flowering is brief, but it feeds the first pollinators and produces seeds that sustain future growth. The authors use this image to describe their own working model: model releases will be "the visible flowers," while the training recipes and the research they enable are what they hope will keep nourishing the scientific ecosystem long afterward.

This announcement is best read for what it is: a founding statement and a call to action — hiring, funding, compute — not a report of results. The text does not yet include published recipes, released models, staffing numbers, a dated roadmap or technical details on which architectures or model scales they plan to tackle first; all of that remains, for now, a promise. Its news value lies in the diagnosis it offers about the progressive closing of post-training research in the AI industry, in the attempt to give that diagnosis a concrete institutional structure backed by already-identified funders, and in the fact that it comes signed by a figure with a recognizable track record in the open AI space, which lends the initiative some initial weight within that community.