What kind of layer 3s make sense?

Vitalik Buterin argues that stacking rollups on rollups does not produce large scalability gains, because rollup data cannot be compressed twice, though layer 3s can serve other purposes such as customized functionality. He cites Starkware's layer 3 framework, which proposes separate layers for purposes like privacy, application-specific environments and validiums. Buterin notes that tokens can move between two layer 2s committed to the same layer 1 without L1 transactions, though with a fraud-proof-window delay on deposits. He also says that a ZK rollup inside a ZK rollup would reduce fixed proof costs to about 8,000 L1 gas, though he presents an alternative batching approach as well.