<div dir="ltr"><div>I have been seeing (for months) that the motivations to love or hate LLM are very diverse. And that I cannot predict my friends and colleagues position just based on what I know from them.</div><div><br></div><div>So far I am very concerned about the accuracy and correctness of the LLM answers in many domains. Why are we so flexible with their mistakes? They are not with mines!</div><div>If my professional answers were as bad as the LLM I tested, I would be probably unemployed. But we treat LLMs as a stupid student that we have to correct all the time.</div><div>Sometimes I have the feeling that it is like the student that read a lot of books, but didn't understand the concepts. Later in the exam, they are just throwing some answers. If they are lucky, it works. But they are not always lucky.</div><div><br></div><div>Another thing that worries me is they are breaking the way we humas estimate the correctness of an answer. When somebody speaks very well our native language, it usually means that person is very educated, and that brings more knowledge and accuracy (sometimes it is not the case... and they become offenders or politicians).</div><div>LLMs speak very well. That lowers our warnings, and we accept any answer. They look correct; but we all know they can be very wrong. We humans have been improving this way to guess the other human feedback for millennia. Now we are failing applying it to LLMs,... because they are not humans.</div><div><br></div><div>That together with their inability of answering "I don't know" is very dangerous. (I also don't like to be adulated as they do, but that is more a personal opinion).</div><div><br></div><div>Also there are many outdated answers. Once I asked about PyQGIS, and after some iterations (because it was not working) it said "oh, you are right, this is not compatible with QGIS3" What? were you using QGIS2 documentation? I gave up.</div><div><br></div><div>My experiences with Gemini Pro have been disappointing. I had to correct it again and again. Even in topics where it should be much better than me, like CSS.</div><div><br></div><div>The friends and colleagues that use it more are confirming the brain rot it is already producing to them. It is a drug: offered cheap to create an addiction. Then you cannot leave it. I know cases of people that use Clause to re-run the QA tests every single time. Not to create the script, but to run the tests! Every time!</div><div><br></div><div>Finally, I am waiting to see a raise in prices. The companies that migrated massively to use LLMs will have three options: a) pay the huge increase in the price b) come back to "human work", c) bankrupt. (maybe also a combination of a and b). Some of them destroyed so much their human workforce that option b is not possible. Perfect work for the "drug dealer"</div><div>Why that hurry to use LLMs for everything? Companies (I mean users, not the AI providers) can just wait and see. There is always time to get into the game, but quitting it can be very expensive.</div><div><br></div><div>Cheers.</div><div><br></div><div>PS do you remember how Radium (the element) was used for everything when it was discovered? Drinks, clothes, watches... radium everywhere. Now we know its risks and we use it only when it is needed, by trained people. LLM is used by everybody, without any understanding.</div><div><br></div><div>PPS Why are some managers insisting that much to use it? When somebody insists that much, I can only be suspicious. It never was for the better. Never.</div></div><br><div class="gmail_quote gmail_quote_container"><div dir="ltr" class="gmail_attr">On Sat, 27 Jun 2026 at 13:46, Even Rouault via Discuss <<a href="mailto:discuss@lists.osgeo.org">discuss@lists.osgeo.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">my turn... Various unordered thoughts<br>
<br>
- My personal experience as a "practitioner" with LLM has been <br>
essentially limited to chat bots, and experimenting with LLM code <br>
reviews. The today's online chatbots are quite good when you're an <br>
expert in the field where you query them and can have a critical eye to <br>
assess their suggestion. They might be practical you need to do <br>
throw-away coding for things you have no clue about. I guess that's the <br>
end of the positive part of this message :-)<br>
<br>
- I tried a couple months ago making a few "open" models work on my <br>
computer, a laptop from 2020 with 32 GB RAM and 4 GB graphics memory. <br>
Just testing chatbot experience, that "worked" in the sense, it had the <br>
performance of a ChatGPT of 3 years ago that is good at making you <br>
laugh, and at a very slow pace.<br>
<br>
- In a project like GDAL, bumping something as the version of the C++ <br>
standard generally lead to some heated conversation about if it is <br>
really needed, if it is the good time to do so, if we couldn't defer by <br>
a few more years because you know someone still needs to build a modern <br>
GDAL with a 10+ years compiler. Technical conservatism is kind in the <br>
DNA of the project, so adoption of breakthrough technology is not <br>
something in our habits<br>
<br>
- We (GDAL) had a burst of low value/low quality vibe coded <br>
contributions towards the beginning of the year, with some strong <br>
correlation with Google Summer of Code preparation (which the project <br>
doesn't participate at. We issued a first version of an AI tool use <br>
policy that was mostly inspired by the LLVM one whose philosophy is <br>
"there must be a human in the loop". That wasn't very effective. We had <br>
to considerably harden it: <br>
<a href="https://gdal.org/en/stable/community/ai_tool_policy.html" rel="noreferrer" target="_blank">https://gdal.org/en/stable/community/ai_tool_policy.html</a> . I strongly <br>
suspect that some PRs we receive don't follow those guidelines, but hard <br>
to prove and humanly messy to deal with.<br>
<br>
- We also add a burst of LLM based vulnerability reports, but quite <br>
modest regarding the size of GDAL and its huge vulnerability surface. <br>
Likely reason: GDAL has been enrolled in OSSFuzz since many years and we <br>
already have fixed hundreds of issues. I'm not sure if we're completely <br>
done on the front of LLM based vulnerability reports though.<br>
<br>
- The point of failures between LLM based contributions and human based <br>
ones are different. It is very difficult to debunk errors in LLM based <br>
contributions, because you get comments all over the place which tend to <br>
make your critical eye asleep.<br>
<br>
- One recurring trait I've observed is overly verbose tests compared to <br>
what a human do. Like 3 times more verbose. It is likely humans don't <br>
write enough tests, but it is not like more tests is always better. They <br>
also add to technical debt. Anectodal evidence of that: many years ago <br>
(well before AI was a thing), one contributor nearly spent one year <br>
refactoring the whole GDAL test suite from its home-made framework to <br>
pytest (I guess some could say: an agent would do that in a few <br>
hours/days nowadays)<br>
<br>
- AI bot scrapping has visible effects on our infrastructure. <br>
ReadTheDocs which hosts <a href="http://gdal.org" rel="noreferrer" target="_blank">gdal.org</a> and <a href="http://proj.org" rel="noreferrer" target="_blank">proj.org</a> documentation was (and <br>
apparently still is) victim of DDOS, and had to turn on CloudFare human <br>
check tests to access pages, for a whole range of IP address, which <br>
includes mine. I assume AI bots must be behind that. End result: some of <br>
our resources, like JSON schemas, that are supposed to be accessible <br>
through a simple curl request are no longer accessible by a portion of <br>
users.<br>
<br>
- The amount of contributions by some people who adopt agentic AI is <br>
strongly demultiplied, at least in quantity. On a couple examples <br>
recently seen at looking at their github activity, I've seen 10 or 16 <br>
times bumps between 2024 and 2026. I guess that this could seen as a <br>
win, but that can strongly modify project dynamics if some developers <br>
adopt it at scale, and others don't<br>
<br>
- I'm *very* concerned by the environment impacts of those technologies, <br>
at least those which are used in practice today and are in the hands of <br>
a few big players. They are clearly not the only responsible of all our <br>
sins in that domain, but it is again one more thing humanity didn't <br>
fundamentally need and adds to its long basket of things that are <br>
heading towards the wrong direction, and are likely to be very hard to <br>
compensate.<br>
<br>
- What could be an acceptable AI for me ? Maybe something like:<br>
* that has been trained on material whose licensing allows to do so <br>
and/or material whose authors have explicitly approved used for LLM <br>
training (some explicit tag in /robots or equivalent)<br>
* where scrapping is done at a reasonable rate, not effectively <br>
DDOS'ing websites<br>
* you can make run inference / agents standalone on your own <br>
average computer with normal hardware specs<br>
* whose environmental impact is similar to normal computer uses<br>
<br>
Even<br>
<br>
<br>
-- <br>
<a href="http://www.spatialys.com" rel="noreferrer" target="_blank">http://www.spatialys.com</a><br>
My software is free, but my time generally not.<br>
LLMs contribute to global warming and brain rot<br>
<br>
_______________________________________________<br>
Discuss mailing list<br>
<a href="mailto:Discuss@lists.osgeo.org" target="_blank">Discuss@lists.osgeo.org</a><br>
<a href="https://lists.osgeo.org/mailman/listinfo/discuss" rel="noreferrer" target="_blank">https://lists.osgeo.org/mailman/listinfo/discuss</a><br>
</blockquote></div>