Wonder if this tracks with the fact that when you're actively thinking time feels like it slows down but when you're passively absorbing time feels faster.

Thusly letting llms do the 'heavy lifting' for five minutes while you dither feels like 1 minute of actual coding. But then you need to also spend 2 minutes fixing up the llm output resulting in 8 minutes wall time. Time prompting + generation + fixing. Vs time just coding directly.

I misread this as "[a]ll coding tools make developers shower," and briefly imagined how clean that subset of people could become.

I should go back to sleep.

It's one study. It's pretty sturdy in terms of methodology but it hasn't been peer reviewed as far as I can tell. They also only looked at established software projects, not anything new. So this is a narrow scope and it doesn't prove that so-called AI cannot enhance productivity at all. It just indicates that pro devs can be fooled into thinking they are better off with it when they are not. But I feel like that's hardly news in these mad times.