AI and efficiency

In three linesOpenAI releases analysis showing that since 2012, compute required to train a neural network to equivalent ImageNet classification performance halves every 16 months. Training to AlexNet-level performance now requires 44x less compute than 2012 — far exceeding Moore's Law's predicted 11x improvement. Algorithmic progress outpaces classical hardware efficiency for heavily-invested AI tasks.

Summary generated by Claude — human-verified