The per-ref serialization demo makes the point better than the OpenAI post itself. Nothing got faster, the work just stopped waiting in line. That is the whole lesson in one sentence.

The causal consistency tradeoff is the part worth sitting with. Dropping global ordering sounds scary until you ask who actually consumes it. Almost every consumer cares about one ref, one document, one tenant. Paying the global ordering tax to preserve an invariant nobody queries is one of the most common silent costs in production systems.

Also appreciated the caveat at the end. A single lock with 40 users is correct engineering, not technical debt.