People are naturally attracted to the idea that “first something is expensive, then it gets cheaper” — as if AI is a single thing of constant quality, and when it gets cheaper, we’ll use fewer chips to train it. But what’s important is the scaling curve: when it shifts, we simply traverse it faster, because the value of what’s at the end of the curve is so high.
