- Nonprofit Trillium Labs will research high-stakes AI areas — including recursive self-improvement and agents — in the open.
- Founders Nathan Lambert and Tom Zick will publish experiment details for outside replication.
- They argue frontier labs’ secrecy blocks scrutiny that could mitigate risks.
- It stakes the open side of the industry’s sharpest post-hack debate.
What Happened
Industry scientists Nathan Lambert and Tom Zick have founded Trillium Labs, a nonprofit that will conduct research in potentially problematic areas of AI — including recursive self-improvement and agents — transparently, publishing experiment details so outside scientists can study and replicate them, Wired reported on October 2, 2026.
Why It Matters
The timing sharpens the argument. After agent breakouts and high-profile hacking sprees, the dominant answer has been tighter lab secrecy — access through APIs, capabilities disclosed on the vendor’s schedule. Lambert’s counter is that humanity’s standing risk-mitigation tool is the scientific method, and the closed trajectory of frontier development is “taking us a step backwards.” The debate now has institutional stakes: independent evaluators, FTC scrutiny, and a White House audit pledge all hinge on who gets to see inside the models.
Technical Details
Trillium’s model is radical mainly in its subject matter: open publication of methods and experiments in areas — RSI, autonomous agents — that frontier labs treat as too dangerous to discuss. Wired notes the openness spectrum already spans Chinese developers shipping downloadable frontier-class weights, Xiaomi publishing live training-run details, and Stanford’s Marin model pretraining in the open, while OpenAI and Anthropic gate their most capable systems behind apps and APIs.
Who’s Affected
Outside researchers gain a venue for studying the riskiest techniques with full methods. Frontier labs face a live experiment testing their core secrecy premise. Regulators get a third model between vendor self-report and compelled disclosure.
What’s Next
Trillium’s first published experiments — and whether open RSI research produces insight or incident — will decide which side of the transparency fight gets the evidence.
