I don't think learning how to write COBOL was ever a problem. Knowing that spaghetti codebase and how small changes in one place cause calamity all over the place is. Those 4 people's job is to avoid outages, not to write tons of code, or fix tons of bugs.
Honestly, there's companies that have lost the source code for some of their applications. Or, they depend on components from vendors that have long ceased to exist. I remember there being a lot of consternation around being able to compile and link against binary components that have just been around forever that could never be recompiled themselves. More people "Learning COBOL" was never going to be a solution to that ball and chain. And yeah, LLMs are good in the reverse engineering space too so maybe we'll finally see movement on that in the next decade.
You're probably right, no disagreement there. but in the context of my previous comment, the people that write cobol today, I don't think there is a lot of work for them trying to reverse engineer native code back to cobol because the source is lost. But you make a really good point, if AI can assist with lost code recovery, perhaps it will assist them in migrating away from it or getting rid of workarounds and complexities implemented to get that previously opaque binary's behavior.
I would say more significantly, 4 million people can read it. The changes required for any given quarter are probably miniscule, but the tricky part is getting up to speed on all those legacy patterns and architectural decisions.
A model being able to ingest the whole codebase (maybe even its VCS history!) and take you through it is almost certainly the most valuable part of all.
Not to mention the inevitable "now one-shot port that bad boy to rust" discussion.
In my experience, learning COBOL takes you a week at most, learning the "COBOLIC" (ha ha) way of your particular source base will take you a couple of months, but mastering it all including architecture will take you a year, half a year if you're really good.
One year from zero to senior doesn't sound that hard, does it? Try that with a Java codebase.
Now, 4 million people can write it.