The counter argument is that now you can skip boilerplate code and focus on the overall design and the few points that brainpower is really needed.
The amount of visualizations that i have made after chat gpt was released has increased exponentially. I loath looking the documentation again and again to make a slightly non standard graph. Now all of the friction is gone! Graphs and visuals are everywhere in my code!
> focus on [...] the few points that brainpower is really needed
The person you're responding to is talking about it from an educational perspective though. If your fundamentals aren't solid, you won't know that exponentially smoothed reservoir sampling backed by a splay tree is optimal for your problem, and ChatGPT has no clue either. Trying things, struggling, and failing is crucial to efficient learning.
Not to mention, you need enough brain power or expertise to know when it's bullshitting you. Just today it was telling me that a packed array was better than my proposed solution, confidently explaining why, and not once saying anything correct. No prompt changes could fix it (whether restarting or replying), and anyone who tried to use less brainpower there would be up a creek when their solution sucked.
Mind you, I use LLMs a lot, including for code-adjacent tasks and occasionally for code itself. It's a neat tool. It has its place though, and it must be used correctly.
The amount of visualizations that i have made after chat gpt was released has increased exponentially. I loath looking the documentation again and again to make a slightly non standard graph. Now all of the friction is gone! Graphs and visuals are everywhere in my code!