AI agent swarm reportedly beats state of the art on Karpathy’s NanoChat benchmark in three days
The team describes a custom research setup built on a graph database, with each iteration learning from mistakes.
TLDR
A team claims its swarm of AI agents beat the state of the art on Karpathy’s NanoChat benchmark in three days. It describes building a custom research setup on top of a graph database, with each iteration learning from mistakes to improve the research process. The post says the team wrote more than 15,000 entries.
Combined views
1.9K
1 Source, first seen 15d ago