Isn't that what is already happening? I'd argue that LLMs will get feedback-loops and other, specialized LLMs for crosschecking and permanent storage like we store information in our brain it can access (isn't all that vectoring very similar to how neurons work - now we might just need some upscaling, minituarization, cheaper hardware and specialiced sections just like in the human brain. So, I guess that "needs human oversight, context and special
knowledge"-part will go away. AI is still in it's infancy, we got the principles of LLMs down, but most LLMs are very generally trained and agentic AI is still new, Niw there's the long dark road of optimization and connection of different parts ahead. So, yes, programmers will become AI-babysitters, but I wouldn't build a career on that.