On 2025-06-05, Eli Schwartz wrote:
On 6/5/25 12:16 PM, Mark Knecht wrote:
<hehe>
I sense a certain overriding fear about the future...
</hehe>
I'm not worried about the future, I'm worried about today, and today's
users seeking help and getting led astray.
To each his own. I agree with you in general - don't trust AI with
anything important. If you prefer then 'trust but verify'.
I don't consider AI to provide value enhancement for me at all, for much
the same reason as described at
https://infosec.exchange/@david_chisnall/113690087142854474
It saves time on writing code. It costs *more* time in debugging subtle
edge cases. This is a steep disadvantage for a technology that is full
of questions about legal liability!
Of course, your mileage may vary... if it's only for personal use you
may not care about legal liability, and if you couldn't write it
yourself at all then maybe the debugging is worth it.
Edge cases, corner cases, implementation details and non-portable
features are reasons why you'd want to *avoid* "GenAI" for shell
scripting.
I'd say a bit of shell scripting is being aware of what the conditions
to be careful with are. That sounds precisely like the kind of thing at
least some public GenAIs might fail at. Possibly even if you stick to
GNU bash and Linux (which shouldn't be a problem unless you plan on your scripts being used on other systems).
I can't forget the moment when I got a GenAI telling me how to use
killall to kill processes *by name* in Solaris [1]. That said, maybe
that public model was defective, given it claimed the white side of a
Space Shuttle Orbiter goes *down* on top of a carrier aircraft...
[1]
https://social.sdf.org/@njsg/113130981279894435
--
Nuno Silva
--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)