Introducing Trillium Labs

Fostering the open science of frontier AI.

Website | Announcement on X / LinkedIn | Hiring Form

Modern AI grew out of a scientific commons. Decades of research conducted across universities and industry laboratories, shared through papers, code, and public benchmarks have led us to the AI systems frontier labs are releasing today. For much of that time advances were happening through open research. Peer review and open methods let a diverse community challenge assumptions, reproduce results, and take them in directions that their original authors never anticipated. That community also trained the researchers now building the frontier. AI reached this point through science as we have practiced it: knowledge accumulating across institutions, with people able to inspect and build on each other’s work.

That pattern continued into the early generative AI boom. Papers on the transformer architecture and reinforcement learning from human feedback gave the broader community access to techniques that became central to commercial systems. Attention is All You Need (the paper introducing the transformer architecture as we know it) came from Google, OpenAI’s InstructGPT paper described methodology for applying RLHF at the level of specific algorithmic design choices. Researchers outside the company could investigate these methods, adapt them, and help develop the field.

Even as issues with these techniques have become central, not just for the field, but the public at large, the published science behind them has significantly thinned out. As post-training has become more important to reasoning and agentic capabilities, access to complete recipes is scarce. Researchers can observe the results of increasingly consequential training decisions while having limited ability to investigate the decisions themselves.

Frontier labs turning the research spigot down to a trickle is hardly surprising. The cost of transparency has risen with the commercial value of that knowledge. And the sums at stake are extraordinary. The Wall Street Journal reports that Anthropic’s proposed IPO could raise up to $100 billion — an offering that would exceed the roughly $78 billion raised by Saudi Aramco, Alibaba, and SoftBank Corp’s IPOs combined. A method that makes an agent more reliable can also make it a better product; publishing it gives competitors a piece of the knowledge underpinning those valuations.

Meanwhile, reproducing a serious training program requires enough compute and engineering that academic researchers cannot easily fill in the missing details themselves. The commercial stakes increase the incentive to keep methods private, while the expense of independently studying them puts that work beyond the reach of much of the scientific community.

These companies could be entering one of the largest periods of corporate growth we have seen. They are also developing a foundational technology whose effects will extend far beyond their customers and shareholders. Aligning these systems, securing them, and making their benefits available across society are difficult problems. For centuries, science has helped us address problems of this scale by making methods explicit, testing claims independently, and accumulating evidence across institutions. The same process that brought AI to the frontier is essential to understanding what we are building there.

We believe these problems require a scientific community with the resources to investigate them independently. A handful of closed research programs cannot provide the diversity of questions, methods, and perspectives that a technology this consequential needs.

That is why we’re founding Trillium Labs. We want to sustain the scientific process that made the frontier possible, beginning with post-training. We will build fully open post-training recipes, including the data, code, evaluations, and intermediate checkpoints that let other researchers study how model behavior develops. Producing those resources is expensive. Once they exist, a much broader community can use them to test interventions, investigate failures, and adapt models to problems the original team would never have thought to pursue.

We want to build on the foundations of fully-open work done in the community, from the Allen Institute for AI, EleutherAI, OpenAthena, Nvidia, and Hugging Face, but they alone are not enough.

The nonprofit structure makes producing that infrastructure our central commitment. We can invest in controlled experiments, document failed runs, and investigate how a training intervention affects behavior across subsequent stages and release all of it without worrying about protecting our IP. We can publish findings that complicate our own assumptions. Our measure of success is how much independent research those resources make possible.

We’ve begun with support from Halcyon Futures and Schmidt Sciences, and we’re fundraising from a diverse coalition to sustain this work. We want funders with different perspectives on AI’s trajectory to support a shared scientific foundation. They can disagree about which outcomes are most likely or which risks matter most while agreeing that more researchers should have the resources to test those claims.

Our name comes from trilliums, spring ephemerals that bloom before the forest canopy fills in. Their flowers are visible for only a short time, but they provide resources for early pollinators and produce seeds that support future growth.

That is the inspiration behind our lab. We want to seed the scientific commons around post-training while there is still room to shape how the field develops. Our model releases will be the visible blooms. The training recipes and research they enable are what we hope will keep nourishing the ecosystem long afterward. This will grow into many cycles of addressing the biggest open questions facing frontier AI with open science.

We’re hiring, we’re fundraising, and we’re searching for compute. Please get in touch!

Nathan Lambert & Tom Zick