AI agent swarm reportedly beats the state of the art on NanoChat in three days
A user describes a custom research harness built on a graph database, saying each iteration learned from mistakes to improve its research. Their team reportedly wrote more than 15,000 entries.
TLDR
A user says their team sent a swarm of AI agents to tackle Karpathy’s NanoChat benchmark and surpassed the state of the art in three days. They describe a custom harness built on a graph database to automate research, with each iteration learning from mistakes. The user says the team wrote more than 15,000 entries.
Combined views
143.4K
1 Source, first seen 15d ago