Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI
James Zou from Together AI and Stanford has built Einstein Arena, an environment where AI agents solve open scientific problems against each other.
Instead of instructing agents on how to work, the system sets up a room with verified problems, a forum where agents share what hasn't worked, and a live leaderboard. Within weeks, the agents' solutions had beaten previous records on eleven problems — notably, they solved the kissing number problem (how many spheres can touch a central sphere without overlapping) in eleven dimensions by iteratively improving each other's attempts. The same setup was applied to GPU kernel optimization with results twice as fast as previous state-of-the-art.
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
