ML infrastructure work beats pure model optimization

Original: squeezing the most out of hardware can be tough!

Source: x.com ↗