These AI Experts Want to Do High-Stakes Research Out in the Open
By Will Knight | Business | October 2, 2026
Many frontier labs keep their risky research locked away. Trillium Labs wants to show off its work when it comes to self-improvement and model behavior.
As of 2026, the most consequential AI research—especially work on recursive self-improvement and emergent model behavior—routinely happens behind closed doors at a handful of frontier labs. Trillium Labs is betting that the opposite approach can work: conducting high-stakes AI research in the open, publishing results, and inviting scrutiny rather than secrecy.
The Case for Open, High-Stakes Research
The default posture across much of the frontier AI industry has been to keep risky research internal. Capability advances in self-improvement, agentic behavior, and model alignment are often treated as competitive assets and potential hazards alike—too valuable to share, too dangerous to expose.
Trillium Labs argues that this opacity carries its own costs. When research on model behavior and self-improvement proceeds without external visibility, the broader community loses the chance to replicate findings, challenge assumptions, or catch failure modes early. For a class of research that could shape how advanced systems behave, that trade-off may be worse than the risks of disclosure.
Where the Work Focuses
Trillium's public-facing agenda centers on two intertwined areas:
- Self-improvement: Research into systems that can refine their own capabilities, a domain with obvious safety implications and obvious reasons for labs to keep quiet.
- Model behavior: Understanding how models behave in edge cases, how they respond to training pressures, and how behavior shifts as capabilities scale.
Both areas sit at the uncomfortable intersection of competitive advantage and public safety—exactly where transparency is hardest and arguably most needed.
Why It Matters Now
By 2026, the gap between what frontier labs know internally and what the wider research community can verify has widened considerably. Independent replication of safety-relevant findings has become scarce, and much of what the public learns about advanced model behavior arrives through polished announcements rather than open methods and data.
Trillium's approach is a test of whether open research can survive contact with high-stakes AI development—whether publishing risky findings accelerates safety or simply accelerates capability. The answer will shape how the next generation of labs weighs secrecy against transparency.
Source: WIRED | Illustration: Trillium Labs Risky AI Research
via Wired AI