Then, too, hand-crafted assembly still beats compiled output, if you consider the only upper limit on quality. The gap in quality is much, much smaller, and the gap in time and cost is much, much bigger, though, so it's almost never worth the trade-off.
Let's say that AI-generated software eventually (or already, if you like) produces better quality software than people can produce, given the same time and cost restrictions. You can still spend more time and cost to get a better result, or accept a worse result to save time and cost. Or take more time to get a good result but more cheaply using off-peak pricing (tokens or power, take your pick). Or throw more parallel agents at the problem to get a good result faster, but at higher cost.
Pick two. Always.
I hit a corner case yesterday in which assembly would give me a better result than compiled code. Emulating an eZ80 ALU on an x86_64 platform means doing a simple operation like add then taking something like 26 more instructions to sort out the flags. Or you can just use ‘mahf’ to load the x86 processor’s own flag results into a register and use them almost verbatim.
It’s still not worth doing, though, since I can easily reach performance targets without.
And beyond this you're also speaking of something different. You're talking about taking some output assembly and then modifying it in 'negligible' (as a percent of the whole) ways, as opposed to writing assembly oneself. For a program of minimal complexity, not only would it would require an unreasonable amount of effort, but you'd also invariably write inefficient assembly relative to a compiler. Compilers produce code that is better, faster, and cheaper.
---
Shifting outside the abstract, let's just do an example. Imagine you have a processor running at 3 billion cycles/s and you have a program that takes up to 1 million single cycle ops/s. And then some ultra clever human tunes a bit of the assembly and we get it down to 0.95m cycles per second. In reality you're probably not going to be able to tune your way to a generalizable 5% boost, but even if you do in this scenario - what have you achieved? Literally nothing. Of course this argument, taken to extremes, is why it now takes more processing power to run a text editor than it did to put a man on the Moon, but we're not talking about extremes.