Distinguish between inference scaling and "larger tasks use more compute"

A viral analysis argues we must distinguish between two reasons AI inference costs are skyrocketing: 1) LLMs are tackling larger, more complex tasks (good), and 2) they are using more compute per task for the same output (bad). The latter, true 'inference scaling,' threatens progress because compute costs can't be amortized like training. If most recent gains come from wasteful scaling, not efficiency, AI's economic viability hits a wall.