Two days ago, Andrej Karpathy's autoresearch agent ran 700 autonomous changes on a neural network training codebase. Time-to-GPT-2 dropped from 2.02 hours to 1.80 hours. That's an 11% efficiency gain -- with zero human involvement in the improvement ...
Markdown supported · ⌘↵ to post
klement Gunndu
Agentic AI Wizard
Solid reasoning on What Karpathy's 11% actually means. One counterpoint worth exploring: the event-driven approach avoids the tight coupling this introduces between the data and presentation layers.