Re: Path prefix to src and targets

[David Deutsch]

Hi Paul,

Well, I was mostly testing the waters, so thanks for clarifying.

I have checked the competence of GPT4 for make many times before I 
decided to test it on mailing list questions. I have always found it to 
be surprisingly good.

In this case, I checked the response, too and found it to be appropriate 
and helpful. (And do let me know if I'm missing something here!)

The thing that wasn't quite up to spec was the part about stems - the 
response did not go into detail about how the stem can turn out to be 
somewhat less than intuitive. It wasn't necessarily wrong (as far as I 
can tell), but it lacked detail. When pressed for more details, I think 
it does a good job:

https://chat.openai.com/share/0f7d6168-3126-4844-9fac-99b704e6eec7


This might be getting a little too meta, but I'm only a casual observer 
of this mailing list and I've seen basic questions go unanswered a few 
times. I suppose that's not surprising. make is, unfortunately, seen by 
many as a dated relic and not a core competency. So in my view, with 
audience dwindling, this might become more of an issue in the future.

I've toyed with the idea of building a fresh introduction to make for 
newbies, explaining the core tech and how it's important to understand 
make as a graph processing tool where its simplicity is not a sign of 
limitation, but of decades of honing down a very sharp edge that is as 
valid now as it used to be. (I often joke that many software 
architectures - build systems especially - contain an ad hoc, 
informally-specified, bug-ridden, slow implementation of half of make.) 
I've not yet found the time to do that, but I still might, in the 
future. Seeing how well GPT4 fares with explaining make to new users (or 
even to people like myself who have used it for quite a bit longer), I 
think there might be a part of that project that is about building a 
virtual expert that developers could use to hone their own make skills, 
like a make "tutor". It would be hard, I think, to find a human tutor 
like that.

My point is - perhaps it's even getting increasingly harder. There is an 
element of "erosion of competency" in the user base of many trusted 
tools and that's a real shame, but - such is life. LLMs might not be the 
silver bullet here, as you can never "simulate" a good community, but 
perhaps they could be helpful in aiding those who are still around with 
the workload of keeping it alive?

-David


On 30.05.23 14:20, Paul Smith wrote:
> On Mon, 2023-05-29 at 06:06 +0200, David Deutsch wrote:
> > I hope I'm not breaking any etiquette rules with this, but I was
> > interested in seeing how GPT4 would respond to this request
> My attitude about GTP answers is that they're OK to post, AS LONG AS
> you (the person who posts the answer) fully understands and agrees with
> everything in the response.
>
> If you don't know for sure that the response is accurate from your own
> knowledge, please don't post it (or else, verify it yourself before
> posting it).
>
> I have no problems with people using GPT (etc.) to help formulate an
> answer for them, since not everyone is proficient at writing clearly in
> English.  But, I've seen many responses posted to StackOverflow (where
> GPT responses are not allowed, but the response was clearly generated)
> that look reasonable and correct at the start, or at a high level, but
> are very wrong in the details.
>
> The last thing we want to do is disseminate plausible-sounding, but
> incorrect, information.
>
> So TL;DR: if you're using GPT et.al. to help you write a more clear
> version of information you know to be correct, that's fine.  If you're
> just copying text back and forth between the mailing list and GPT
> without fully understanding, please don't and let someone who knows the
> answer respond instead.
>
> Cheers!