Parameter count is only a coarse measure of capacity. Through a least-squares toy problem and three LLM examples — LoRA rank, Transformer depth, and MoE experts — we show that larger can even hurt, and argue that what matters is how much capacity training actually exploits.

Drawing on biological and urban scaling, Braess's paradox, and recent LLM observations, we conjecture that selfish learning — neurons crowding into easy feature directions at the expense of global coordination — may become the next bottleneck for scale.