[The Duke]

"but, when i was in the computer science program at UC (ug and grad) we had two grades for programs: A for it works perfectly and basically F for anything else. admittedly that was back in the 70’s. different time, perhaps different standards. so for me, if i have to fix someone’s or something’s program for them, it’s always an F. ergo my statement, ‘AI can’t code.’"

And, when I was writing code for the Space Shuttle Mission Simulator (SMS) and NORAD we were similarly picky. But, when storing business data that can be re-created ... not so much.

The code to control AZ/EL for SETI telescopes was also a bit critical. Folks got unreasonably upset when they had to replace gear boxes due to excessive acceleration and jerk.

Go figure.

[MV=PY]

Yes, code is more important if it can bend metal (developed early Z80 motion control for printing presses).

[dadfly]

yes, i guess that would be a bit upsetting for the MEs :).

based on your previous comment, i was wondering if a test AI could be trained to write and run simple unit tests, given general set of requirements, and then learn from how and when a system/emulation was failing to write and run some integration testing. that could be useful and take some of the boredom out. it could certainly keep trying to exercise as much of the code as possible, maybe compile stats on intermittent failures, and maybe even start detecting failure patterns.