My Three Wishes for "Open" AI

About this event

Huan Sun

As AI grows more capable, autonomous, and potentially self-improving, scientific access and human intellectual agency matter more than ever. We must make AI more “open” and a better subject of science. To me, openness goes beyond access to model weights or APIs: it means being able to independently build and experiment with AI systems, rigorously evaluate what they can—and should not—do, and investigate the mechanisms underlying frontier AI’s capabilities and failures.

In this talk, I will share three lines of research toward this vision: building capable agents with fully open training recipes; testing agent capabilities and unintended behaviors; and understanding the potential and limitations of transformer architectures, particularly looped transformers. Overall, these efforts represent some steps toward a more open and rigorous science of frontier AI.

This talk is sponsored by AI-ACCESS .