AI prioritizes getting the result right
while ignoring whether the basic premises make sense
Building skyscrapers on a fragile foundation
with no steel frame running through the whole structure
Each new floor is just another iteration
and it grows more and more rickety in the wind and rain
so you keep adding more patch-and-paste props
to hold up a system on the brink
Worse still
you don't know how many hidden defects and blind spots remain
or when it will suddenly collapse
For example
the coding agent I've been building behaved oddly on Vercel serverless
On the surface, the code set a state variable in single-process memory
but across HTTP requests in serverless you can't guarantee the same process
so AI's fix was to move that state to a centralized DB/Redis
plus optimistic locking with CAS for concurrent writes
My gut said it shouldn't be this complicated
After digging through it
the real issue was bad flow design causing concurrent writes
That can be avoided with separation of responsibilities
and simple orchestration so there's only one place that writes
https://t.co/XA87IzCBKp
Skip to content
AI prioritizes getting the result right
From X · @therealfreddyx
Thoughts
mew～
AI prioritizes getting the result right
while ignoring whether the basic premises make sense
Building skyscrapers on a fragile foundation
with no steel frame running through the whole structure
Each new floor is just another iteration
and it grows more and more rickety in the wind and rain
so you keep adding more patch-and-paste props
to hold up a system on the brink
Worse still
you don't know how many hidden defects and blind spots remain
or when it will suddenly collapse
For example
the coding agent I've been building behaved oddly on Vercel serverless
On the surface, the code set a state variable in single-process memory
but across HTTP requests in serverless you can't guarantee the same process
so AI's fix was to move that state to a centralized DB/Redis
plus optimistic locking with CAS for concurrent writes
My gut said it shouldn't be this complicated
After digging through it
the real issue was bad flow design causing concurrent writes
That can be avoided with separation of responsibilities
and simple orchestration so there's only one place that writes
https://t.co/XA87IzCBKp
while ignoring whether the basic premises make sense
Building skyscrapers on a fragile foundation
with no steel frame running through the whole structure
Each new floor is just another iteration
and it grows more and more rickety in the wind and rain
so you keep adding more patch-and-paste props
to hold up a system on the brink
Worse still
you don't know how many hidden defects and blind spots remain
or when it will suddenly collapse
For example
the coding agent I've been building behaved oddly on Vercel serverless
On the surface, the code set a state variable in single-process memory
but across HTTP requests in serverless you can't guarantee the same process
so AI's fix was to move that state to a centralized DB/Redis
plus optimistic locking with CAS for concurrent writes
My gut said it shouldn't be this complicated
After digging through it
the real issue was bad flow design causing concurrent writes
That can be avoided with separation of responsibilities
and simple orchestration so there's only one place that writes
https://t.co/XA87IzCBKp