On 5/13/2026 10:11 PM, Janis Papanagnou wrote:

AIs are not better, they just pick samples from their random
knowledge base.

I thought that too at one point, but it's mostly not true.

Open Google.com and click on AI Mode and ask this question:

does Google Gemini regurgitate programming code

The answer will surprise you.

It may be complete trash what they provide.

That could be an interesting challenge: get the worst AI-code possible.

I "beat" AI not too long ago, by asking for the program (not SQL) to
sort results by 3 groups.

* group 1
* group 2 within group 1
* group 3 within group 2 within group 1

The code it delivered for the first 2 worked, but the 3rd just hung. I
don't remember if it was C or python.

AIs have also no focus. - They can't do programming, they
associate from the question some (good or bad, functional or
hazardous) code. That code is not different to what you call
"facts". They are as bad in choosing appropriate programs as
they are presenting other "facts"; quality is purely random.

I know someone that developed a 'vibe-coded' tcl/tk-based Usenet reader
in an hour or so. It would've taken probably a week by hand.

Of course it required refinements, but still a huge time-saver.

In my discussions with chatGPT, I have encountered errors
very often. You really have to be skeptical when using it.

That is sound advice. - I'd even say, get an experienced
programmer and leave AI's products alone. The AI stuff is
nice to play with and if you don't need to produce reliable
software in a larger project's scale. It may provide you,
if you're inexperienced in software development, with hints,
but never use the products where lives or big money depends
on.

+1



--- PyGate Linux v1.5.19
* Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)

On 7/26/2026 3:21 PM, DFS wrote:

On 5/13/2026 10:11 PM, Janis Papanagnou wrote:

AIs are not better, they just pick samples from their random
knowledge base.

I thought that too at one point, but it's mostly not true.

Open Google.com and click on AI Mode and ask this question:

does Google Gemini regurgitate programming code

The answer will surprise you.

They do seem to generate code dynamically, but how smart they are is
pretty variable. A lot of time they are prone to generate very verbose boilerplate-laden code.

If the goal is efficient code and/or optimizing for a niche target, not
so great. Or, they will over-complicate things or hallucinate the
ability to use features that they can't use.


I have seen cases where it (well, mostly Grok in my case) falls on its
face if you ask it to stick with C90 style and not to use anything
beyond the C standard library. It kept doing non-C90 things, like
declaring variables in places variables can't go in C90.

Well, or asking it to write code to perform large decimal arithmetic but
work with 32-bit values in a form where each represents a value from 000000000..999999999, ...


They have gotten better at least, like was impressed a while back when
Grok was smart enough to figure out how to go about filling in a tricky
finite state machine (as a semi-difficult challenge).

Previously, they were failing at tasks like, say:
Write a C function to take a color in RGB555 form, and map it to the
closest match in the 16-color CGA RGBI palette.

But, last I checked, they are up to this task (in an earlier test, this apparently stumped the AI, which took a long time and then gave
something completely unrelated, like how to invoke an image converter
API in Python, IIRC...).

I am like, I would have accepted something like, say:
int RGB555ToRGBI(uint16_t rgb)
{
static const int tab[64]={0, 0, 1, 1, ... , 15};
int ix;
ix=((rgb>>9)&0x30)|((rgb>>6)&0x0C)|((rgb>>3)&0x03);
return tab[ix];
}
Or similar...


But, as noted:
ChatGPT:
Not really used;
To do much of anything with it costs money.
Gemini:
Mostly gives matter-of-fact responses;
Sometimes goes off on weird / conspiracy theory level tangents.
Grok:
Often a little more clever than Gemini;
Also more prone to obvious hallucinations;
Keeps trying to claim to have done things it couldn't have done;
...
Copilot:
Seems to mostly give reasonable responses, but not used as much.

There are some of the models one can run locally, but they are painfully
slow and not usually smart enough to be worth the wait.

It may be complete trash what they provide.

That could be an interesting challenge: get the worst AI-code possible.

I "beat" AI not too long ago, by asking for the program (not SQL) to
sort results by 3 groups.

* group 1
* group 2 within group 1
* group 3 within group 2 within group 1

The code it delivered for the first 2 worked, but the 3rd just hung.ÿ I don't remember if it was C or python.

One of the great limiting things is that the chat-box forms can't deal
with anything larger than what one can upload in a few text files, and responses that can fit into a chat response.

And, if the text file is all that large, they break...


Or, if they set "agentic AI" loose on something, they often just end up
with a mess (or wrecked codebase, or the AI messes up their computer, ...).


If you want information about using an OS API or similar, they can help
with this at least.

AIs have also no focus. - They can't do programming, they
associate from the question some (good or bad, functional or
hazardous) code. That code is not different to what you call
"facts". They are as bad in choosing appropriate programs as
they are presenting other "facts"; quality is purely random.

I know someone that developed a 'vibe-coded' tcl/tk-based Usenet reader
in an hour or so.ÿ It would've taken probably a week by hand.

Of course it required refinements, but still a huge time-saver.

In my discussions with chatGPT, I have encountered errors
very often. You really have to be skeptical when using it.

That is sound advice. - I'd even say, get an experienced
programmer and leave AI's products alone. The AI stuff is
nice to play with and if you don't need to produce reliable
software in a larger project's scale. It may provide you,
if you're inexperienced in software development, with hints,
but never use the products where lives or big money depends
on.

+1

--- PyGate Linux v1.5.19
* Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)

On 08/05/2026 23:45, fir wrote:

i must say i was probably never sorry i was born in late 70-ties becouse thanx to that i live in 90-ties but now im a bit sorry becouse those 15
years (2000-2015) spend on coding witchout ai seem now terrible path
compared to new ...

If we hadn't all done that then chatgpt would have had no C coding data
for it to induce relations among.

Although, why would we need it to generate C code when it's internal requirements representation, NN and final determiner is an adequate
program itself.

It would do a more valuable job as a query-tool for its own such state -
we'll soon be so unskilled that having the C-code isn't explanatory for
us and we'll have to load the C-code into it to ask whether it's okay.
Since the creation of C-code from its state is itself just a query
anyway, it is an illusion that the C-code it makes is meaningfully useful.

--
Tristan Wibberley

The message body is Copyright (C) 2026 Tristan Wibberley except
citations and quotations noted. All Rights Reserved except that you may,
of course, cite it academically giving credit to me, distribute it
verbatim as part of a usenet system or its archives, and use it to
promote my greatness and general superiority without misrepresentation
of my opinions other than my opinion of my greatness and general
superiority which you _may_ misrepresent. You definitely MAY NOT train
any production AI system with it but you may train experimental AI that
will only be used for evaluation of the AI methods it implements.

--- PyGate Linux v1.5.19
* Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)