Re: [ELPA] New Package: minuet-ai.el

[Eli Zaretskii]

> Date: Sat, 15 Mar 2025 11:44:31 +0100
> Cc: rms@gnu.org, philipk@posteo.net, dev@milanglacier.com,
>       monnier@iro.umontreal.ca, stefankangas@gmail.com,
>       emacs-devel@gnu.org
> From:  <tomas@tuxteam.de>
> 
> On Sat, Mar 15, 2025 at 12:15:47PM +0200, Eli Zaretskii wrote:
> 
> [...]
> 
> > (Btw, AFAIU this package is specially tuned to use code-generating
> > models, and those are a far cry from the ones you call "bullshit
> > generators", which produce human-readable text with "expert
> > opinions".)
> 
> They are exactly the same LLM architectures, used on code.

Not necessarily: some of them use a stage of supervised learning after
the initial unsupervised one, and that second stage tunes them to be
much more useful and correct.

> And they "hallucinate" (their term) or "bullshit" (our term) in very
> similar ways.

If a model produces invalid code, a compiler or interpreter will
quickly discover and flag that, and unit testing (which is a must
regardless whether the code was written by a human or not) should
definitely find that, so the chances of a good model to "hallucinate"
are lower in this case.

> Their business model is producing code which "looks" correct,
> just as their cousins do with lanuage. Arguably, this is even slightly
> more dangerous.

When did you last use such a model?  And what "hallucinations" of code
did you get when you did?

In any case, this is becoming off-topic here; if you want to continue
this discussion, please move it to emacs-tangents.