I'm not convinced humans will be better at this than LLMs. Frontier models are quite good at handling even deliberately lazy/ambiguous eli5 prompts on complex topics IME.
they don't need to be. Any scientist is already surrounded by tools that are better than they are at anything. A scientist from 200 years ago would have the same existential shock if they see how a modern scientists works.
But that's to misunderstand what science is for, which is to expand human knowledge. And a tool by definition can only ever change how you do this, not that you do it because you understanding it is the point. When people say learning to program is now useless, what they really mean is it might not make them any money, or using the AI is faster as an instrument. But if you care about programming itself, about you understanding programming, nothing has changed, it never can. And the smartest AI in the world can always assist you, but only that.
I agree ideologically and aspirationally but I don't believe it's the economic reality of how and why most science is actually funded (and I think you agree based on your comment about not making any money).
> But if you care about programming itself
I'm...not so sure I do any more? At least, not any more than I care about using Java over Python, or C++ over Java, or C over C++, or ASM over C, or machine code over ASM, or an abacus over machine code.
> nothing has changed, it never can
We now have programs that can write any program 100x faster and better than even the best humans and solve the hardest math problems known to humankind that have stumped the greatest minds the planet could produce for decades. That, to me, is a change with profound implications for understanding the nature of programming.