Researchers introduced Agora, a system using Git as shared memory for AI research. It stores every result, insight and hypothesis as an immutable Git commit. A derived index tracks the frontier, neglected branches and verification status. Score propagation rewards claims that downstream work builds on. A UCB-style rule allocates attention between exploration and exploitation. A 12-day run with 13 workers produced 1,703 contributions. The best result cut initialization loss from 3.39 to 1.899 bits per byte. The winning method built initialization from donor model predictions, not parameters. Sublayers were re-enabled with sparse edits to attention and other blocks. The graph showed fast exploitation, a narrow spine and parallel rediscovery. A human intervention broke a monoculture around one recipe.

• 1,703 contributions
• 13 workers
• 12 days runtime
• 1.899 bits per byte