An autonomous agent optimized Andrej Karpathy's training pipeline overnight, discovering gains that years of manual tweaking missed. The experiment shows that human intuition can lag behind algorithmic search, making researchers themselves the new bottleneck.
The Signal
It highlights the growing role of self‑optimizing agents in accelerating AI research and reducing human effort.