But does that change the end result? Finding a compiler or SQL bug doesn't mean you yourself can learn enough to fix it. I don't see any reason why AI would be inherently incapable of also concluding that there's a bug in an underlying layer beyond its ability to fix

That doesn't mean AI can do or replace everything but what fraction of software engineering work requires that final frontier?

They can, by putting what they just "learned" into the context window. Claude Code does this without (my) prompting from time to time, adding to its CLAUDE.md things that it has learned about the project or my preferences. Currently this is limited to literally writing it down, but as context windows grow and models continue training on their own usage, it's not clear to me how that will significantly differ from an ability to "learn information they currently do not have".