Fantastic take. Matches my own experience 100%. Well written, and funny, too!
What's not mentioned is that these "weeks of architectural decisions" that must happen in minutes are exhausting, just as they are when we spread them over a long period. Perhaps more so!
I'm finding that being forced into these huge, multi-module level decisions many times per day wears out my brain and if I don't stop and get away from it, the LLM will just happily continue to output stuff that isn't going the correct "big" direction.
So while it's still true for me that this method gets to the finished product faster, it's not true that you get back 100% of that time as dev time. You don't. Maybe housecleaning, reading, TV, something else time? But not heavy brain work time.
Stwerner · 15h ago
Yeah that’s definitely been my experience too. It’s just as exhausting, compressed into a much shorter amount to time. I’ve gotten more endurance as I keep pushing, kind of reminds me how pair programming felt the first few months of doing it all day.
threecheese · 16h ago
We need to have more conversations about the Nth-order impacts of this new-normal “age of unlimited code”. Which of our patterns do we throw away when “intra-human communication across time” is no longer as important a constraint within our codebases, when developers don’t need to use code to understand how it works?
Where are new module boundaries? Do we even worry about duplication and modularization? Do we ship data models as documentation and test cases, and let codebases implement them however they like, whenever they like? My mind swims.
What's not mentioned is that these "weeks of architectural decisions" that must happen in minutes are exhausting, just as they are when we spread them over a long period. Perhaps more so!
I'm finding that being forced into these huge, multi-module level decisions many times per day wears out my brain and if I don't stop and get away from it, the LLM will just happily continue to output stuff that isn't going the correct "big" direction.
So while it's still true for me that this method gets to the finished product faster, it's not true that you get back 100% of that time as dev time. You don't. Maybe housecleaning, reading, TV, something else time? But not heavy brain work time.
Where are new module boundaries? Do we even worry about duplication and modularization? Do we ship data models as documentation and test cases, and let codebases implement them however they like, whenever they like? My mind swims.