I agree with most points made by the author, however the cognitive impact depends on how you use LLMs. Here's an example: I build so many Phoenix / Elixir applications and one of the things I really hate is the amount of boilerplate code I need to write for every project. It is exactly the same (eg. Setting up vite, removing existing mix based node integrations) and modifying scaffolded code to adhere to my design. I already know CSS inside out and Tailwind CSS too. So, by offloading the task of rewiring existing scaffolded CRUD templates, I would argue the opposite - I am saving precious cognitive load from doing repetitive data entry work.

That's why I have a very clear personal policy of always being the architect and instructing LLMs what AND how a task must be done as opposed to blindly just trusting them with everything - that is how you lose your ability to think.