Nvidia unveils Agora, using Git as shared memory for AI research agents
Posts describing Agora say 13 agents worked for roughly 12 days without assigned tasks or a central planner, closing 62% of the performance gap to a trained GPT-2 on a weight-transfer challenge.
TLDR
Two posts describe Nvidia’s Agora as a system where research agents build on one another’s work through a shared Git project history. One account says the task was to initialize a new model using knowledge from 141 existing models, without training data or training runs. The posts report 1,703 contributions from 13 agents over roughly 12 days. The more detailed account also reports 165 checks of other agents’ work, none of which found a failure.
Combined views
—
2 Sources, first seen ago
