Subj : Re: Writing with LLMs
To   : Nightfox
From : tenser
Date : Thu Sep 18 2025 11:49 pm

On 17 Sep 2025 at 09:02a, Nightfox pondered and said...
 
 Ni>   Re: Re: Writing with LLMs
 Ni>   By: Bob Worm to Dmxrob on Wed Sep 17 2025 08:23 am
 Ni> 
 Ni>  BW> It worries me how much I see people depending on ChatGPT - people in 
 Ni>  BW> people I sit next to on the train, university professors... I feel li
 Ni>  BW> brain is rusting up enough from age without speeding it up by outsour
 Ni>  BW> my thought to a machine as well.
 Ni> 
 Ni> I'm a software engineer, and recently my manager at work asked me if
 Ni> I've tried using Copilot to write any code. Maybe it's just to try to
 Ni> justify the company buying Copilot licenses..  I don't really want to
 Ni> rely on AI to write code though, at least not fully.  I've also seen
 Ni> AI-generated code be very wrong, or sometimes it would give the same as
 Ni> something I'd find with a Google search.

I've been playing with it recently (well, Claude, not ChatGPT).
I've found it about 2/3 miss and 1/3 hit.

Recent tasks include using it as a glorified search engine,
asking for specific points about the use of a particular
library I wanted to program against, for example.  It was...
ok for that.  I asked it to generate an example program, and
it was able to do so, but that program was deficient in some
ways.  When I asked it to explain the rationale behind some
of the choices, it hallucinated a bunch of nonsense.

Another time, I asked it to explain what, I'm pretty sure, is
a bug in the GNU assembler; specifically, I made it do a bunch
of analysis that I could have done myself, but which would
have been tedious and time consuming.  It did an adequate job
of pinpointing the section of code I wanted it to, but
alternated confidently between asserting that the observed
behavior was intended or a bug.

In both cases, my conclusion was that, had I not already had
decent sense of what was going on in both cases, I'd have spent
a _lot_ of time chasing wild geese.

Automated generation of boilerplate code is all well and good,
but even for that, I'm not really seeing the value-add.  My
real desire is to use these tools as a research assistant, and
for that, they still fall well short of the mark.  Like a bad
intern that speaks with the confidence of a pretty big ego on
one hand, but the sycophantic obsequiousness of an ass-kisser
waiting to stab you in the back.

--- Mystic BBS v1.12 A48 (Linux/64)
 * Origin: Agency BBS | Dunedin, New Zealand | agency.bbs.nz (21:1/101)

.