According to WIRED, a new AI research organization called Trillium Labs is taking an unconventional approach to some of the field’s riskiest work by conducting it in the open rather than behind closed doors. While most frontier AI laboratories keep their research on potentially dangerous topics under wraps, Trillium Labs wants to publicly showcase its work on areas including self-improvement and model behavior.
The approach represents a departure from the secretive practices common among leading AI companies, which typically restrict access to research they consider high-stakes or potentially risky. According to WIRED, Trillium Labs believes that transparency in these critical areas can benefit the broader AI research community. The organization’s focus areas of self-improvement, where AI systems could potentially enhance their own capabilities, and model behavior, which examines how AI systems act and respond, are considered particularly sensitive topics in AI safety discussions.