[Please do not mail me a copy of your followup]

Lynn McGuire <[email protected]> spake the secret code <vurv36$340f4$[email protected]> thusly:

I call bull hockey pucks.

Have you tried any generative AI assistants? They're remarkably good
at generating boiler plate in the small and the large. However, there
are caveats (as always).

I think this article is a good summary of where we are today: <https://addyo.substack.com/p/the-70-problem-hard-truths-about>

The other thing to consider is that generative AI assistants are only
going to get better from here.

If you haven't yet tried one, try GitHub Copilot in Visual Studio, but
be aware that you'll quickly burn through your budget of 2,000 code
completions per month on the free acount. This is because it's
counting all the suggestions that are being made against that budget
and not just the ones that you accept. I burned through the 2K
suggestions in a weekend of coding.

-- Richard

--
"The Direct3D Graphics Pipeline" free book <http://tinyurl.com/d3d-pipeline>
The Terminals Wiki <http://terminals-wiki.org>
The Computer Graphics Museum <http://computergraphicsmuseum.org>
Legalize Adulthood! (my blog) <http://legalizeadulthood.wordpress.com>

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

On 30/04/2025 23:30, Chris M. Thomasson wrote:

On 4/30/2025 10:38 AM, Richard wrote:

[Please do not mail me a copy of your followup]

Lynn McGuire <[email protected]> spake the secret code
<vurv36$340f4$[email protected]> thusly:

I call bull hockey pucks.

Have you tried any generative AI assistants?  They're remarkably good
at generating boiler plate in the small and the large.  However, there
are caveats (as always).

I think this article is a good summary of where we are today:
<https://addyo.substack.com/p/the-70-problem-hard-truths-about>

The other thing to consider is that generative AI assistants are only
going to get better from here.

If you haven't yet tried one, try GitHub Copilot in Visual Studio, but
be aware that you'll quickly burn through your budget of 2,000 code
completions per month on the free acount.  This is because it's
counting all the suggestions that are being made against that budget
and not just the ones that you accept.  I burned through the 2K
suggestions in a weekend of coding.

Last year I asked AI to write some lock-free code... It tried, but got
the memory barriers all messed up. Now, I did not try to correct it
because, well, why should I. Wow. Bug city.

AI (currently) is useless for that kind of thing. It can write code
that /looks/ correct, based on common examples - but for code where you
need to think through the consequences of the sequencing, or where the
time and effort is in planing the code or in analysing its correctness,
AI hasn't a chance.

It can be okay for simple, repetitive boiler plate code or as a slightly
more advanced version of code refactoring tools. And if it makes some
tedious parts of the job go faster, that's fair enough. (I'd be much
happier if AI could do some of the documentation, or writing time lists,
or attending customer meetings...)

It doesn't help that AI learns from existing code samples, and a lot of
that is crap code or simple demo code.

I first remember reading about how software development will be done by computers and most programmers will be out of a job some 35 years ago.
I remain unconvinced.

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

On Thu, 1 May 2025 11:09:39 +0200
David Brown <[email protected]> wibbled:

I first remember reading about how software development will be done by >computers and most programmers will be out of a job some 35 years ago.
I remain unconvinced.

Indeed. Boiler plate code interface or GUI code isn't an issue even for 4GL tools, the actual program logic is another matter. To tell an AI the fine details of what you want the program to do you'd need to write such a comprehensive spec you might just as well right the code yourself anyway.

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

On Fri, 2 May 2025 11:15:44 +0200
Bonita Montero <[email protected]> wibbled:

Am 01.05.2025 um 11:39 schrieb [email protected]:

Indeed. Boiler plate code interface or GUI code isn't an issue even for 4GL >> tools, the actual program logic is another matter. To tell an AI the fine
details of what you want the program to do you'd need to write such a
comprehensive spec you might just as well right the code yourself anyway.

However, code generators aren't needed for boilerplate code because
this code is trivial anyway. However, I've presented my code to ChatGPT >several times, and ChatGPT claimed to simplify it, but the ssimplifica-
tions contained errors. I mean, if an AI is already so overwhelmed with
less than 100 lines of code, things can only get worse with more complex >code.

My experience with it writing code isn't great either though it seems better with plain C rather than C++.

However my young daughter tried to get chatgpt to explain some basic science to her. It got archimedes principal the wrong way around saying things float better in less dense liquids. After that I told her not to bother asking it anything and to use google instead.

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

On Fri, 2 May 2025 15:51:42 -0000 (UTC), wrote:

On Fri, 2 May 2025 11:15:44 +0200
Bonita Montero wibbled:

Am 01.05.2025 um 11:39 schrieb :

Indeed. Boiler plate code interface or GUI code isn't an issue even for 4GL >>> tools, the actual program logic is another matter. To tell an AI the fine >>> details of what you want the program to do you'd need to write such a
comprehensive spec you might just as well right the code yourself anyway. >>

However, code generators aren't needed for boilerplate code because
this code is trivial anyway. However, I've presented my code to ChatGPT >>several times, and ChatGPT claimed to simplify it, but the ssimplifica- >>tions contained errors. I mean, if an AI is already so overwhelmed with >>less than 100 lines of code, things can only get worse with more complex >>code.

My experience with it writing code isn't great either though it seems better >with plain C rather than C++.

However my young daughter tried to get chatgpt to explain some basic science to
her. It got archimedes principal the wrong way around saying things float >better in less dense liquids. After that I told her not to bother asking it >anything and to use google instead.

search the right data as answer it is not far from a AI

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

On 4/30/25 17:30, Chris M. Thomasson wrote:

On 4/30/2025 10:38 AM, Richard wrote:

[Please do not mail me a copy of your followup]

Lynn McGuire <[email protected]> spake the secret code
<vurv36$340f4$[email protected]> thusly:

I call bull hockey pucks.

I can believe that 30% figure. It just explains so much. :)

Have you tried any generative AI assistants?  They're remarkably good
at generating boiler plate in the small and the large.  However, there
are caveats (as always).

I think this article is a good summary of where we are today:
<https://addyo.substack.com/p/the-70-problem-hard-truths-about>

The other thing to consider is that generative AI assistants are only
going to get better from here.

If you haven't yet tried one, try GitHub Copilot in Visual Studio, but
be aware that you'll quickly burn through your budget of 2,000 code
completions per month on the free acount.  This is because it's
counting all the suggestions that are being made against that budget
and not just the ones that you accept.  I burned through the 2K
suggestions in a weekend of coding.

Last year I asked AI to write some lock-free code... It tried, but got
the memory barriers all messed up. Now, I did not try to correct it
because, well, why should I. Wow. Bug city.

Nobody should expect AI to create "correct" code, whatever that means.
It's a meaningless term. It's just used to say "there's no bugs as far
as we know".

You could prove specific assertions or properties about a program.
But even there there is not a lot language support to give an AI
anything to reason with. Maybe with c++ contracts for fairly trivial assertions. Definitely not for concurrent code. C++ doesn't have it.
Rust doesn't have it, it just has the "unsafe" label which means that
you have to trust that somebody knows what they are doing.

Joe Seigh

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)

[Please do not mail me a copy of your followup]

jseigh <[email protected]> spake the secret code <vvvkni$1rrq7$[email protected]> thusly:

On 4/30/25 17:30, Chris M. Thomasson wrote:

On 4/30/2025 10:38 AM, Richard wrote:

[Please do not mail me a copy of your followup]

Lynn McGuire <[email protected]> spake the secret code
<vurv36$340f4$[email protected]> thusly:

I call bull hockey pucks.

I can believe that 30% figure. It just explains so much. :)

As someone who observed the inside workings of MS for 10 years in the
early 2000s, most people have no idea what goes on inside MS. They
don't understand that the things they don't like about MS software
aren't the fault of the engineers, but instead are the fault of
"program managers" and the career ladder inside MS.

Nobody should expect AI to create "correct" code, whatever that means.
[...]

From the way you're talking it sounds like you haven't actually tried
any of these assistants. It doesn't work the way you think it does.
--
"The Direct3D Graphics Pipeline" free book <http://tinyurl.com/d3d-pipeline>
The Terminals Wiki <http://terminals-wiki.org>
The Computer Graphics Museum <http://computergraphicsmuseum.org>
Legalize Adulthood! (my blog) <http://legalizeadulthood.wordpress.com>

--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)