> Ever since I got my hands on my first computer as a kid, I've been outsourcing parts of my brain to computing so that I can focus on more interesting things. I no longer have to remember phone numbers, I no longer have to carry a paper notepad, my bookshelf full of reference books that constantly needed to be refreshed became a Google search away instead. Intellisense/code completion meant I didn't have to waste time memorizing every specific syntax and keyword. Hell, IDEs have been generating code for a long time. I was using Visual Studio to automatically generate model classes from my database schema for as long as I can remember, and even generating CRUD pages.

I absolutely agree with you, but I do think there's a difference in kind between a deterministic automation you can learn to use and get better at, and a semi-random coding agent.

The thing I'm really struggling with is that unlike e.g. code completion, there doesn't seem to be a clear class of tasks that LLMs are good at vs bad at. So until the LLMs can do everything, how do I keep myself in the loop enough that I'll have the requisite knowledge to step in when the LLM fails?

You mention how technology means we no longer have to remember phone numbers. But what if all digital contact lists had a very low chance of randomly deleting individual contacts over time? Do you keep memorizing phone numbers? I'm not sure!