I literally do packaging for a living and you are misunderstanding my point. Most people just take a binary and run it. There's no analysis of the assembly code. You might profile it and bench it after the fact but no one is sitting there looking at the assembly line by line unless there's a very very good reason and frankly LLMs are better at that type of investigative work. I know because I've been investigating some curious 1 in 100,000 segfaults recently and guess what? It took an LLM to build a tool to let us even hit that bug because it was basically impossible to do by hand and no one in the before times would have sat down to write the tool cause we would not have time so we would have just accepted that 1 in 100,000 requests are segfaulting. At least now I can actually fix the problem.

What's the reliability of compilers this day? How likely for a bug to be in your code and not in the compiler? I think it's close to 99.99...

So when you have a bug and a core dump, you can quickly load it in debugger, see the stack frame and then theorize a model for the bug to happen. If after verifying the source and having complete confidence that it's good, then you start looking at the assembly, most likely while single stepping with the debugger. But you rarely get to that point, because 99.99... it's your code.

That reliability is what AI tooling is lacking. It's exhausting monitoring the output because errors can be as simple as a minus character or the wrong comparison operator.