I suspect the difference lies in people who think in code versus people who translate their thoughts into code

I think in code and programming concepts when I am writing software . I don't really know how to explain that, but I don't often feel like there is any friction between my thoughts and the code I make

I think that many coders do not have this. They have an extra "translation" step that introduces friction into their workflow

I don't experience this friction, so LLM coding introduces new friction that I don't like

They do experience this friction, so LLM coding doesn't introduce new friction for them, it may transform their previous friction into a new form that is easier for them to navigate

I don't know. Maybe I'm talking out my ass, this is just a random theory based on no real evidence

Another datapoint of one.

It's not really that I think in code. It's more like code is as much as a language to me as writing english, or holding the pencil to draw something. I got a change request, or I read something from the docs, and the the mental concept I have realign itself to the new knowledge. Getting code out is always effortless. There's no difference between

  return the list of names of all the books that have the fantasy tag

And

  return books.filter(b -> b.tags.contains('fantasy')).map(b -> b.name)

If I can describe something, I can code it.

This is it for me as well. It comes down to which is quicker to write. If I need an exact solution or implementation, then I just write it, because the code is significantly more efficient to describe the problem in than in plain english. An exception being visual things like plots, if we just care about a way to get to a plot, vague english is fine and likely less effort than writing code.

i.e. function to calculate things, just write it; want to plot some data, vibe it. I can see immediatly if its not quite right and in a line can ask for the change i needed, zero cognitive effort required to fix issue, and hey presto, we have what we want.

i think this stems from the fact natural language is generally quite inefficent for logical statements / arguments / instructions but can gather visual ideas far more efficiently than code can.

Yes, this is exactly what I mean when I say "thinking in code". I feel it isn't accurate either, but it was the best way I had to describe it

More or less just that there is little to no friction for me to think a thought and write the code compared to "translating" the thought into code