Hi there, I'm Cam Allen.
I'm a postdoc studying AI with Stuart Russell at the Center for Human-Compatible Artificial Intelligence (CHAI) at UC Berkeley. In spring 2024, I was also a lecturer for CS 188, Berkeley's upper-division artificial intelligence course. Previously, I completed my PhD at Brown University, advised by George Konidaris.
My research goal is to steer artificial intelligence towards human flourishing. This requires studying both the computations that drive intelligent behavior and the alignment protocols that ensure AI systems do what we want. In both areas, my work addresses a fundamental challenge that has existed since the earliest days of AI: identifying an appropriate conceptual frame.
A frame is a mental representation that defines the scope and structure of the task to be solved. The frame problem—identifying what is relevant and how to represent it—is essential for building AI systems that do anything, let alone what we want. For any fixed frame simple enough to enable efficient decision making, there exists a task that cannot be solved without the information that frame ignores. And since what we (humans) want changes over time and we don’t always even know what it is that we want, our AI systems must be flexible enough to adapt. Until we develop a principled way for AI systems to construct their own frames, and modify them as necessary, open-ended tasks like “ensuring human flourishing” will remain forever out of reach.
My research program studies both the theory and practice of building AI systems that construct their own frames, along with techniques for inspecting the frames of existing systems and aligning AI frames with those of humans.
- Frame Construction. How should agents form abstract representations of what they know? How can agents construct appropriate conceptual models of their decision problem? How can agents detect when their frame is missing information? How can agents build frames that flexibly adjust to new tasks?
- Frame Inspection. What new tools do we need to understand the frames of AI systems? What representations and conceptual models do AI systems already learn implicitly, and can we identify similar mental structures in humans? How can an agent notice its own frame?
- Frame Alignment. How might we align the reference frames of different decision-making agents? How can we learn human-interpretable state representations—and skills that precisely manipulate them? Do certain frames lead to self-awareness and/or meaning? How can we build agents that help us refine our preferences and discover what we truly want?