Continuation of `Having "blessed" 3rd party libraries may make D more popular` DIP thread

[H. S. Teoh]

> On Thursday, 10 July 2025 at 12:55:42 UTC, GrimMaple wrote:
[...]
> > I guess it's just a big gray area, because nowadays it's not
> > _really_ regulated by the law, so it can go anywhere from here. But
> > to stay on topic, I must note how LLMs don't write good D code 😅
> > They usually get something super basic done well, but most often I
> > find that the LLM output code needs to be fixed before it can even
> > compile. The other day DeepSeek tried to gaslight me into believing
> > there's no `zip` in `std`, and, once corrected, output utter
> > uncompilable rubbish :)
> > 
> > Usually no such problem with C++ or C#.

That's because the training data probably included a lot more C++/C#
code than D code.

The problem with today's LLMs is basically that they are essentially
interpolation functions, if rather sophisticated ones.  If something was
in their training data, you're likely to get something useful out of it.
But otherwise, it's anybody's guess what comes out.  Probably total
rubbish.  Because LLMs do not actually have a model of comprehension;
they only have a model of what output is more probabilistically likely
given the input.  There are no semantics attached anywhere.


T

-- 
Study gravitation, it's a field with a lot of potential.