Each level of programming abstraction has taken us a step further from the bare metal, as they used to say. In machine code you specify what goes where at the memory/register level. Assembly gives you human-readable mnemonics. C abstracts away direct control, but you still manually allocate memory. Java and C# abstract away memory management but you still declare types. Python and JavaScript abstract away type declarations but you still define variables and program structure. With AIs define your end goal in plain language and then you either have to understand code to edit, or literally depend on the machine to fix everything.

In a sense it's like SQL or MiniZinc: you define the goal, and the engine takes care of how to achieve it.

Or maybe it's like driving: we don't worry about spark advance, or often manual clutches, anymore, but LLMs are like Waymo where your hands aren't even on the steering wheel and all you do is specify the destination, not even the route to get there.

"AI" is not an abstraction like a compiler or a library

it's outsourcing to an unreliable body shop where they barely speak English and the weekly attrition rate is 300%