Thin clients that verify transactions are possible though. For something like Bitcoin you have SPV-proofs that prove chains of headers. You can prove that a transaction was included in the longest chain without having to run a node yourself just by checking proof-of-work merkle trees; Even if the vast majority of users end up running clients that don't verify the whole chain -- cryptographic trust would still be ensured by checking headers. This requires no centralization.

Satoshi wrote about this architecture early on in scaling the blockchain. Ethereum also allows light clients and I think it even has checkpoints that make downloading headers faster. Cryptographic protocols that verify smart contract results could be included in Metamask. I feel like not mentioning this in the essay shows a lack of familiarity with the literature even if he was extremely opened minded (enough to create dapps himself.)

He did make valid observations about third-party trust: OpenSeas and Infura. But in both cases: these protocols can be implemented without centralized architecture. A decentralized alternative to Infura (that provides reliable results to users and easy-to-check attestations) is possible to build. One should also note that in blockchain land the lack of incentives to run a full node is a problem people are working to address. It's actually a perfect illustration of how the blockchain can lead to emergent systems. Some ledgers already have rewards for running full nodes. So yes -- people do want to run full nodes -- they just want to be paid for it.