There is still some room for improvement where trying to introduce LLMs into chaotic codebases and whatnot. LLMs continue to struggle there. Although so do humans, to be fair. Chaos is plain hard. However, in a lot of cases AI is already as good as it can get. "Won't improve much" is a fair take.
Technology has plenty of room to keep marching forward, but the next substantive improvement here will see it no longer be AI. It will become AGI.