I've talked about how new libraries come in with high performance claims and how chasing high performance means making sacrifices here and there. Each little correctness compromise claws back each shortcut and regresses new library back. It plays out every once in a while. Knowing when to stop trying to optimize just because the AI told you so is for the best.
If I told you how many times AI was happy to optimize for the task and throw away constraints, I'd be a trillionaire. AI has no principles, it's there to accomplish the task not guarantee correctness and if it can fudge the numbers, it is happy to do it then call it a mistake/bug later. Out of order execution is obviously going to be faster, but does that mean it is right for the task? Tokio in Rust, for instance is an out of order task scheduler. It is going to be fast when there is no ordering, you have to roll your own sequencer if you want ordered execution. It also brings me back to how Spectre mitigations annihilated Intel chips.
Intel made wild claims about their performance until a correctness compromise demolished their performance and clawed them back. Spectre really destroyed what performance speculative execution offered Intel. Buy new hardware they said. Yeah, oops we didn't prove our claims were correct so now you have to hope the new hardware does what it says and is correct.
Correctness always comes to bite, it's a hard lesson we come to all learn. The chickens came home to roost.