> the genie, used well, accelerates learning.

Ehh... 'used well' is doing some very heavy lifting there. And the incentive structure at 90% of companies does not optimize for 'using it well.'

The incentive is to ship quickly, meaning aim the AI-gun at the codebase for a few hours and grind out a "technically working" solution, with zero large-scale architecture thought and zero built-up knowledge of how the different parts of the application are intended to work together (because there was no "intention"). There will be tests, but they may not be sensible and may be very brittle.

Anyway, deploying a bunch of fresh grads armed not with good mentorship but with the ability to generate thousands of LOC a day is a recipe for accelerating the collapse I usually see in startup codebases about 6-8 years old. This is the point where the list of exceptions to every supposed pattern is longer than the list of things that follow the patterns, and where each bug, when properly pursued, leads to a long chain of past bad decisions, each of which would take days of effort to properly unwind (and that unwinding will also have a branching effect on other things). Also, coincidentally, this is the point where an AI agent is the most useless, because they really don't expect all the bizarre quirks in the codebase.

Am I saying AI is useless? No, it's great for prototyping and getting to PMF, and great in the hands of someone who can read its output with a critical eye, but I wouldn't combine it with inexperienced users who haven't had the opportunity to learn from all the many mistakes I've made over the years.