Subj : Re: Writing with LLMs
To   : Bob Worm
From : tenser
Date : Sat Sep 20 2025 01:44:58

On 18 Sep 2025 at 09:08p, Bob Worm pondered and said...
 
 BW>  > alternated confidently between asserting that the observed
 BW>  > behavior was intended or a bug.
 BW> 
 BW> I heard that this is a result of the way the models are trained.
 BW> Essentially the model gets no "points" for saying that it doesn't know
 BW> the answer to something, whereas a confidently stated answer with some
 BW> grain of truth in it (something you'd call bluffing if a person did it)
 BW> would score at least some points. So you get what's incentivised.

Yup.  That sounds about right.  But the LLM isn't fishing for
points on a true/false quiz; as a tool, that kind of behavior
stinks.

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

.