Re: Help building Pen.el (GPT for emacs)

[Eli Zaretskii]

> From: Philip Kaludercic <philipk@posteo.net>
> Cc: rms@gnu.org,  mullikine@gmail.com,  emacs-tangents@gnu.org,
>   stefan@marxist.se,  bugs@gnu.support
> Date: Sat, 24 Jul 2021 14:49:02 +0000
> 
> > How would one know it's 'long' and not some other data type?
> 
> I am not sure what you mean? "long" makes sense here because Java will
> automatically up-cast any other type to fit.

So you came up with perhaps the single example that exists in the
whole world where the issues I mentioned _might_ not matter, and even
that only under some assumptions.  A feature that aspires to be
generally useful cannot possibly depend on such problematic
assumptions.

> >> since it is mentioned over 6000 times on GitHub (and this method even
> >> has a bug, as the article explains -- but that is a totally different
> >> issue).
> >
> > That's not how AI works: it doesn't just count the number of times
> > something is mentioned.  That usually leads to unsatisfactory results.
> 
> Of course, that would be oversimplifying. At the same time, if the
> training samples have common patterns, a model is more likely to
> reproduce that behaviour.

No, that's not it: a single example repeated in identical form many
times doesn't reinforce the learned pattern.  You need many similar,
but different code samples, and most probably in different languages.