Trillium Labs, a nonprofit founded by Nathan Lambert and Tom Zick, plans to study high-stakes AI behavior and publish experiment details for outside scrutiny and replication, according to WIRED on October 2, 2026. Whether openness makes powerful AI safer—or instead exposes risky capabilities more widely—is the central question, and the lab’s reported launch plans do not settle it.
What Trillium Labs says it wants to study
WIRED reported that the lab’s initial agenda includes post-training, AI agents, reinforcement learning’s effects on model behavior, and recursive self-improvement (RSI). These are plans described at launch, not a record of completed research.
Post-training and reinforcement learning
Post-training is work performed on a large model after its initial build, including fine-tuning. Zick told WIRED that the lab would initially focus on this area. He said, “To understand something like how reinforcement learning scales in post-training, you need significant compute and a lot of careful experimentation,” pointing to the resources required to examine how these techniques affect models.
The question is not only whether reinforcement learning improves a model’s capabilities. It can also shape how the model behaves. WIRED cited sycophancy—models tending to agree with or flatter users—as one behavior of concern.
#1 Best Overall
Agents and recursive self-improvement
Agents are systems that can take actions toward tasks, rather than only respond to a single prompt. The report also says Trillium intends to examine RSI: the possibility that AI systems contribute to research that helps develop new models. The report does not specify the experiments, systems, or release boundaries the lab will use for these topics.
Why the founders argue for openness
Lambert’s argument, as reported by WIRED, is that a closed research process limits scrutiny and contributions from people outside the organizations developing frontier systems. Publishing methods and experiment details could let other researchers examine findings, attempt replication, and propose mitigations. Lambert told WIRED, “The current closed trajectory of frontier AI development is taking us a step backwards.” He also argued that “the scientific method and careful measurement of recent events is the best way to understand new behaviors of AI models.”
The dispute is not simply about whether research should be shared. It is about what should be shared, with whom, and at what level of detail. A paper describing a method, a release of model weights, and access to a hosted model expose different information and capabilities. WIRED’s examples included frontier labs that limit access to apps or APIs, as well as downloadable models that can run on users’ own hardware. It also pointed to Xiaomi’s publication of training-run details and Stanford researchers’ open pretraining of Marin; the report does not establish that these examples have equivalent release terms or degrees of openness.
The safety trade-off: scrutiny versus exposure
Supporters of restricted access argue that powerful capabilities should remain available only to a trusted few. Limiting access may reduce the number of people who can directly use a system, but can also leave outsiders with less visibility into how it was built and how it behaves. The founders’ counterargument is that broader research access can help more people identify risks and develop mitigations.
WIRED mentioned the potential for models to automate software vulnerability discovery and probe systems as part of the safety context. The report did not provide a named quantitative study or measure how prevalent or consequential such activity is, so those examples should be understood as concerns raised in the debate, not measured findings about Trillium’s work.
The competing approaches can be assessed across four practical dimensions:
- Scrutiny and replication: Can independent researchers inspect methods and reproduce results?
- Exposure of capabilities: Could publication or access give more people the ability to use potentially dangerous capabilities?
- Visibility into behavior: What can outsiders learn about how a model was built, tuned, and behaves?
- Resources: Do researchers outside industry have enough compute and other support to reproduce work?
These are useful questions for evaluating a release policy, not evidence that either openness or restriction has already been shown to reduce risk more effectively. WIRED’s account leaves that outcome unresolved.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What was reported about funding and scale
WIRED reported that Trillium had launch funding from Schmidt Sciences, Halcyon Futures, and others, but did not disclose the amount raised. The founders’ reported fundraising target was $40 million to $100 million, and their plan was to spend $30 million on training over the next 18 months. Those figures describe intentions at launch, not confirmed fundraising totals or completed spending.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →What remains unclear about the lab’s approach
The public details in WIRED’s launch report do not answer several questions that will determine what “open” means in practice:
- Which artifacts will be published: papers and experiment protocols, model weights, datasets, code, or some combination?
- Who will decide whether a finding or artifact is too sensitive to release, and what criteria will guide that decision?
- What safeguards will apply to research that could reveal or enable risky capabilities?
- How will outside researchers gain access to the compute and other resources needed for replication?
Until those policies are described, the lab’s stated commitment to publishing experiment details is a meaningful intention, but not enough to judge the risks and benefits of its eventual releases.
Who founded Trillium Labs
WIRED reported that Lambert previously worked at Ai2 and Hugging Face, writes a technical blog, and founded American Truly Open Models. Zick previously worked at Harvard and helped Charles Schwab devise responsible-AI policies. The report says the two met over Zoom during the COVID-19 pandemic while they were UC Berkeley graduate students.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errors
