[conspire] Discussion: Using LLMs the Right Way: 10/1/2025 7pm Eastern Daylight time

[Ivan Sergio Borgonovo]

Sometimes I feel your graciousness a bit performative.

Have you any better argument supporting your thesis other than 
illustrating the history of VCS?
That's not the right way to look authoritative on LLM.

Deirdre just wrote that in other companies LLM can do their work to 
manage Mercurial.
So it seems it's not a problem of the technology itself.

How does it come that a company is able to teach to an LLM to do its job 
managing Mercurial and another not only screw up the training but 
instrument it to make a mess of the work of their developers?

You've 3 separate problems and NONE of them is inherently linked to the 
technology itself.
One is related to training as I suggested
The other is related to instrumentation as I affirmed.
The third is related to the expectations on LLM Deirdre declare she 
doesn't have but then apparently relay on them or doesn't but it is too 
tired to handle it [*].

I don't need any specifically internally trained LLM to get sound 
suggestion for git. I even wrote a tool with NO previous experience on 
gemini's API to automate my git workflow in few hours [**].
Because gemini has a fucking huge corpus on git related problems.
It not only knows the syntax but it has a lot of context about when a 
sequence of commands was used to solve which problem.
It will screw up because LLM machines have no concept of truth or 
coherency that is essential to programming but it will generally 
(statistics plays a big role in the way they work) be more useful than 
when it has to help you using Mercurial (over training is a well known 
problem etc... etc...).

When you're able to automate something... some jobs get redundant.
You free up time for developers and you need less or you don't need some 
competences.

How you're going to invest these savings is a social and economic 
problem. You invest them in better products/services or to keep the 
illusion of ever growing stocks and higher dividends.

Not to mention that yeah AI is hyped exactly for economic and social 
reasons one of them is scaring the shit out of unionizing devs.

What you'll end up is a lot of way less competent devs. Instead of 
hordes of people in sweatshops deciding if an image is a cat or a 
traffic light, you'll probably end up with hordes of people in 
sweatshops deciding if the output of a function is correct.
And this is going to produce sloppy code.

Again LLM are not a revolution, or not a big revolution as advertised.
Everybody by now know how they can screw up and how they can be employed 
in ways they make them obnoxious.
And again for social and economic reasons they are more hyped exactly in 
the places where they are most obnoxious.

It's no wonder that people, precisely because of this, feel a particular 
unease and end up being hostile to the technology itself, and not to 
those who control it.

Coming back to your "I prefer tools that are deterministic"...
I'm a non deterministic developer.
If I was ever able to write deterministic code (I get the same 
requirement and I write the same exact code) it would probably be time 
to abstract it into a library... there is a lot of gray area between 
feeling the need to write a library and writing almost the same code.
Actually that area is HUGE because, trust me, sincerely creative moments 
are rare.

People that advertise themselves has living constantly in creative 
moments are exactly the kind of people that DON'T KNOW SHIT because they 
  never studied and have always delegated true work and think others are 
machines, stupid sheep, so they think their work can't be taken by 
machines... but others people work can. Traits that are common with many 
conspiracists that don't believe in experts...

Now you still will need a human in at least 2 positions in the loop: the 
prompt and the evaluation and integration.
You can produce mediocre code at a fraction of the cost with 
reverse-centaurs or you can produce better code giving power to developers.

Guess what's going to make the difference in what is currently happening?
"Smart" developers hostile to AI[***] for "technical reasons", which 
most of the times are wrong otherwise they would be great fan of AI or 
greed?

Initially textile produced with power looms was sub-par and qualified 
workers got turned into slaves. But automation is way more pervasive now 
and they are not planning to supplant just devs. Or at least that's what 
they shout to the market and to unions.

https://www.businessinsider.com/klarna-reassigns-workers-to-customer-support-after-ai-quality-concerns-2025-9

And this is damaging "consumers" too.

[*] reverse-centaur situation. I probably would prefer to spot a typo or 
a variable used in place of another or a compiler parameter that doesn't 
even exist in LLM generated code from a comfortable apartment in Milan 
than blindly run LLM generated code and check if it "works" in a 
sweatshop in Bangalore but the fact that I think the greatest difference 
is the apartment in Milan says something.
And that's what actually happened before I learnt how to save time and 
not feel pain, but I was free to chose how to use a LLM.
A programmer that use an LLM to speed up its coding has still some 
bargaining power, he still has some valuable know-how, a "tester" is a 
commodity.

[**] we could talk about why I can't fully exploit the tool as I wished 
and why LLM still have a choke point and you could read about what a 
really "open source" LLM should be and talk about copyright and what 
plagiarism in code is... But you can run an LLM on consumer hardware and 
you don't need to destroy an ecosystem for it.
But it is way more fun and *popular* and easy and pleasing to use an LLM 
to be reassured of the cogency of your conspiracy theory and that's 
going to be way more dangerous to your ecosystem than the power and 
water you're going to consume.

[***] Deirdre uhm so your "sloppy use of language" has to be excused but 
mine (Y'all) is incompetence?

BTW the things that mostly resemble AI in terms of looking "intelligent" 
and LLM works on the same basis they just have different specializations 
and most of the techniques are common to both.

Deepblue won't look "intelligent" for today standards and yet it was 
considered AI... LLM can't win a chess match. AlphaZero is neural 
network based.

Techniques like decision trees, SVM, rule-based systems are equally used 
in "traditional" programs and stuff in the chain of neural networks.

The "core" that makes this stuff look "intelligent" is neural 
networks... there is still a lot of things to do.
Filtering, preprocessing and this depends on the type of input any "a 
priori" knowledge of the kind of data you're processing so you can 
"push" some knowledge of the model into the neural network, noise 
filtering, learning optimization bla bla bla...
You don't do stemming on images, you do it on sentences, you don't apply 
rotations to phrases, you do it on images etc... plenty of work, of 
knowledge in tons of fields.

And that's why current state of AI and LLM are far from being a 
revolution but rather an incremental development that blossomed due to 
several factors (faster hardware, huge data for training, economic 
interest to push research in a certain direction...).
The probably most important successive step in AI was the discovery of 
the transformer model that made affordable the computations and opened 
the doors to gen AI. It is revolutionary in terms of impact but somehow 
it was inspired by the availability of a certain kind of hardware and 
some previous work on computer vision.

BTW there are "deterministic" AI and they are generally most of the 
techniques above minus the neural network.

I was playing with neural networks when I was 20ish, I'm 53. I don't 
fucking consider myself an expert nor even a practitioner and I feel 
ashamed to have to brag about it to make a point.

On 10/3/25 7:02 PM, Rick Moen wrote:
> Quoting Ivan Sergio Borgonovo (mail at webthatworks.it):
> 
> [nothing much]
> 
> Fine, you want to concentrate on a core competency of gratuitous
> asshattery when I go out of my way to be gracious, go ahead.  I have
> much better things to spend my time on.
> 
> 
> _______________________________________________
> conspire mailing list
> conspire at linuxmafia.com
> http://linuxmafia.com/mailman/listinfo/conspire

-- 
Ivan Sergio Borgonovo
https://www.webthatworks.it https://www.borgonovo.net