Specialization's hidden network costs

In response to a post by @joshpuckett.

there’s a delicate balance on “build vs buy” or “rebuild vs reuse”

when i joined my team at snap, they were in a strong “copy code if trivial” culture.

same idea with shdcn components.

in some ways, it did improve velocity.

code specialization has much the same tradeoffs as specialization does in the real world. you gain local efficiencies and network costs. what truly scales is when the broader industry standardizes and the network costs of dependency chains goes down.

the cost added through the natural complexity of dependency graphs by sharing specialties is difficult to see.

but now we’re seeing the other side

it’s normal for frontier technology to have this state, but it’s compounded by 2 things:

1. we’re reverting some specialties we had already begun to adopt more broadly

2. and the speed at which AI can produce this customized output which has a high % chance of being slop

I hope we can find a way to land in the middle.