Comment by fidotron

About 15 years ago I was introduced to an environment where approximately a hundred developers spent their lives coaxing a classic style expert system ( https://en.wikipedia.org/wiki/Expert_system ) into controlling a build process to adjust the output for thousands of different output targets. I famously described the whole process as "brain damaging", demonstrated why [1], and got promoted for it.

People that spend their lives trying to get the LLMs to actually write the code will find it initially exhilarating, but in the long run they will hate it, learn nothing, and end up doing something stupid like outputting thousands of different output targets when you only need about 30.

If you use them wisely though they really can act as multipliers. People persist in the madness because of the management dream of making all the humans replaceable.

[1] All that had happened was the devs had learned how to recognize very simple patterns in the esoteric error messages and how to correct them. It was nearly trivial to write a program that outperformed them at this.