Many frontier AI labs keep their riskiest research locked away, fearing potential misuse or unintended consequences. But a new research group, Trillium Labs, is taking a different approach. Founded by researchers who previously worked at DeepMind and OpenAI, Trillium Labs wants to do high-stakes AI research—specifically, research into how to build AI systems that improve themselves—out in the open.
Why Openness?
“We believe that the benefits of openness outweigh the risks,” says Tristan Greene, a cofounder of Trillium Labs. “We want to be transparent about our research so that others can scrutinize it, identify potential problems, and help us develop safer and more beneficial AI systems.”
The decision to pursue open research is a deliberate one. The founders of Trillium Labs believe that secrecy can actually exacerbate the risks associated with AI development. When research is conducted in secret, it’s harder for others to identify and address potential problems. Openness, on the other hand, allows for broader participation and collaboration, which can lead to more robust and reliable AI systems.
What Kind of Research?
Trillium Labs is focusing on two main areas of research: self-improvement and model behavior. Self-improvement refers to the ability of an AI system to improve its own capabilities without human intervention. This is a potentially powerful capability, but it also carries significant risks. If an AI system is not properly aligned with human values, it could pursue goals that are harmful to humans.
Model behavior, on the other hand, refers to the way in which an AI system interacts with the world. This is important because AI systems are increasingly being used in real-world applications, and it’s crucial that they behave in a safe and predictable manner.
The Challenges of Open Research
Open research is not without its challenges. One of the biggest challenges is the risk of misuse. If Trillium Labs publishes its research findings, it’s possible that others could use that information to develop harmful AI systems.
“We’re aware of the risks,” says Greene. “But we believe that the benefits of openness outweigh those risks. We’re committed to working with the broader AI community to develop safeguards and best practices to mitigate the potential for misuse.”
Another challenge is the need for careful planning and execution. Open research requires a high degree of transparency and collaboration, which can be difficult to achieve. Trillium Labs is addressing this challenge by building a strong team of researchers and by actively engaging with the broader AI community.
A New Model for AI Research?
Trillium Labs’ approach to AI research is a bold one. It remains to be seen whether it will be successful. But if it is, it could pave the way for a new model of AI research—one that is more open, transparent, and collaborative.