1) Legality and morality are obviously different and unrelated concepts. More people should understand that.
2) Copyright was the wrong mechanism to use for code from the start, LLMs just exposed the issue. The thing to protect shouldn't be creativity, it should be human work - any kind of work.
The hard part of programming isn't creativity, it's making correct decisions. It's getting the information you need to make them. Figuring out and understanding the problem you're trying to solve, whether it's a complex mathematical problem or a customer's need. And then evaluating solutions until you find the right one. (One constrains being how much time you can spend on it.)
All that work is incredibly valuable but once the solution exists, it's each easier to copy without replicating or even understanding the thought process which led to it. But that thought process took time and effort.
The person who did the work deserved credit and compensation.
And he deserves it transitively, if his work is used to build other works - proportional to his contribution. The hard part is quantifying it, of course. But a lot of people these days benefit from throwing their hands up and saying we can't quantify it exactly so let's make it finders keepers. That's exploitation.
3) Both LLM training and inference are derivative works by any reasonable meaning of those words. If LLMs are not derivative works of the training data then why is so much training data needed? Why don't they just build AI from scratch? Because they can't. They just claim they found a legal loophole to exploit other people's work without consent.
I am still hoping the legal people take time to understand how LLMs work, how other algorithms, such as synonym replacement or c2rust work, decide that calling it "AI" doesn't magically remove copyright and the huge AI companies will be forced to destroy their existing models and train new ones which respect the licenses.