You are not logged in.
Pages: 1
AI is becoming a helpful tool for handling complex ideas and creating new features. I asked the AI where it gets the answers and solutions for the Linux Window Manager and Desktop configurations. This was the reply:
"AI models generate answers and solutions for Linux window managers (WMs) and desktop environments (DEs) by parsing upstream distributions, declarative packaging ecosystems, community dotfiles, and system man pages. Because Linux configuration is predominantly file-based and open-source, AI systems can easily ingest, map, and debug the entire ecosystem."
"The Linux community maintains some of the most comprehensive, text-heavy documentation on the internet, which serves as a primary source for AI training data:
Example -- The Arch Wiki: Universally regarded as the definitive resource for Linux desktop configuration."
I asked this specific question: "Do Fvwm3 answers come from MX Linux forum and Arch Linux forum?"
Yes, a noticeable portion of real-world FVWM3 answers, configurations, and community solutions do come directly from the MX Linux and Arch Linux forums.
#####
This shows the importance of forum discussions, posting solutions, and creating new ideas. Many do not use or have access to forums; AI becomes an important tool for them.
Markku
Offline
AI slop generated much of the post and botnets aggressively sweeping archlinux.org and circumventing measures to prevent that.
Are they freeloaders or did they give something back to archlinux and MX Linux communities ?
EDIT : the above is inspired by a report message from loqs
Last edited by Lone_Wolf (Yesterday 12:51:10)
Disliking systemd intensely, but not satisfied with alternatives so focusing on taming systemd.
clean chroot building not flexible enough ?
Try clean chroot manager by graysky
Offline
Many do not use or have access to forums; AI becomes an important tool for them.
instead of relying on LLMs such users should rather stick to what they do since the "web 2.0" became a think in the late 80s: use thier favorite search engine to find the next place to post thier question
better yet: they should learn to either use thier favorite search engine or the search functions of the mail archives or forums they found in the first step to search for and find thier question already asked and answered - because most issues already came up in one form or another, many questions were already asked the very same way, and users like seth keep repeating answering them
but all of the above is BY FAR better than AI slop!
why: because by the very definition of how LLMs work: uncertainty - or more correct: random chance
ask the same model 10 times the same question you'll get 10 different answers
why are 10 times repeated questions with 10 times repeated answers better: because both come from humans - and with high confidence the questions and replies align
Offline
"web 2.0" became a thing in the late 80s
Did you get that information from an AI? ![]()
Many do not use or have access to forums; AI becomes an important tool for them.
In reality, the forums are and have always been open until the LLM bots started to force changes to this (apparently reddit is gonna close down for unsubscribed users because of that)
Humans are simply lazy, so instead of researching a topic, they will always go for a cheap simple answer where they can blindly follow instructions and turn themselves into most inefficient script interpreter possible - and call that efficient.
The fallacies are
1. your prompt/question can be wrong, https://en.wikipedia.org/wiki/XY_problem
2. the LLM instructions can be wrong but you don're realize that although it might be glaringly obvious, https://bbs.archlinux.org/viewtopic.php … 3#p2314543
3. you're missing a learning opportunity
I'm sure online forums are an important source for LLMs - they're a massive and free information resource.
I'm not sure whether injecting an LLM is a positive development.
Offline
ask the same model 10 times the same question you'll get 10 different answers
My lazy reasoning would be "Its just a "thing" which happens to learn from a large vomited sets of data, rote learning it, matching and predicting next tokens statistically, then getting whipped in RLHF, then some security auditing and refinement, and finally the parent's child is infront of public. Ofc the child would be nervous when a hooman in public asks the same question 10 times
.
"though, speaking of it specifically, A nervous child giving 10 different answers is inconsistent; but each answer comes from a coherent attempt to be helpful(hopefully, from one's POV). An LLM giving 10 different answers is "sampling": it's literally rolling dice over a probability distribution (Im not much in a mood to use technical jargon explaining exactly, google or the AI itself would be your friend here.). The child who says "uhh, 5? no, 7? wait, 3" is uncertain about the answer. The LLM isn't uncertain... it has no stable belief to be uncertain about (IMO | obv, LLM doesn't have hooman-like beliefs or subjective uncertainty. However, it can represent uncertainty through its outputs, and its probability distributions can reflect ambiguity in its predictions.). That's a deeper problem than nervousness. (Though temperature 0 mostly fixes the "10 different answers" reducing variable samples, which still isnt a guarantee for the expected results. FAHH.)
LLMs/AIs makes mistakes, A lot. But hoomans do the same when they are uncertain about something, and seeing some arguments on the internet I feel like "Algorithmic exceptionalism". Hypocrisy-cal hoomans.
But hoomans seem to have some sort of common sense and they get their mistakes quickly and try to correct it (unless they are egoistic and emotionally loaded) whereas these AIs overly hedge and if you got into an argument with them, they will nevertheless try to prove their point until you corner them with facts. Yes, they can also learn at some point but... an end user's time gets wasted trying to prove a fact (some sort of sanity too) (context-window limitations, persistent memory, and subscription restrictions aren't quite the same thing). That's it.
Edit:
/no_I_didn't_answered_specific_question_but_laid_out_my_opinion_to_the_void_ ![]()
Edit2: *sighs, If you got pain reading those, I tried my best to simplify my opinion:
both hoomans and LLMs make mistakes for different reasons. An LLM can generate plausible answers without having verified them, and its tendency to defend incorrect outputs makes troubleshooting frustrating, however, hoomans also confidently spread misinformation, so we shouldn't treat every AI mistake as uniquely unacceptable IMO ![]()
The important distinction is whether the information can be independently verified and whether errors are corrected from/of the hooman, the AI, or the Alien, "does time-to-practicality ratio matches?"
Last edited by 5hridhyan (Yesterday 18:54:41)
Gentoo is where dreams and compile times go to die.
And you'll definitely find me talking to a wall :-)
Offline
LLMs/AIs makes mistakes, A lot. But hoomans do the same when they are uncertain about something
Well… just be a god ¯\_(ツ)_/¯
Offline
Well… just be a god ¯\_(ツ)_/¯
Weeeeeelllll.... from a certain point of view..... there are no differences with the AI and hoomans, because nobody of both are able to think for themselves xD. So I guess carry on ? ![]()
and users like seth keep repeating answering them
but all of the above is BY FAR better than AI slop!
AI won't give me answers with good jokes in the middle :C
Though temperature 0 mostly fixes the "10 different answers" reducing variable samples
It seems that also you need to tweak the gpu, because for parallel batch processing of users the matrix multiplication are processed differently. Higher server loads adds a different "entropy" in the result in comparison to smol loads.
Then it can be a deterministic thing ![]()
str( @soyg ) == str( @potplant ) btw!
Also now with avatar logo included!
Offline
In reality, the forums are and have always been open until the LLM bots started to force changes to this (apparently reddit is gonna close down for unsubscribed users because of that)
As botnets have enough resources (IP address ranges CPU and GPU cycles) to bypass protection mechanisms such as Anubis, OTP, real email address verification and return real time results on protected content I suspect it will have no impact on scrapers wanting to index reddit.
Offline
Well… just be a god ¯\_(ツ)_/¯
Exactly, that's why you are the egyptian god!
You can never be wrong or think wrong like the way hoomans do!! \o/
I get that pointing out hooman imperfection doesn't excuse AI mistakes, nor pointing out current AI structure/infrastructure shouldn't excuse a hooman's mistakes too. Mann those "Things" are literally "wanna be hoomans" controlled by big corps (yeah, including that Ram-o-pholic one too)
Higher server loads adds a different "entropy" in the result in comparison to smol loads.
And to my understanding, different execution orders, GPU kernels, hardware configurations, and numerical precision can produce slightly different results.
Those small differences can occasionally affect which token is selected, particularly when two candidate tokens have similar probabilities...:
-Sampling randomness: The decoding process deliberately samples from a probability distribution.
-Numerical nondeterminism: Different execution paths can produce slightly different numerical results.
-Server load: Workload can affect scheduling and execution paths, but higher load does not automatically introduce randomness into every response
and, different execution paths don't necessarily produce different results, and GPU numerical differences don't necessarily change the selected token. They can, particularly when candidate token probabilities are very close. (Ok, I'm DONE hedging. GRrrrrr.)
So yeah, server-side load... potentially a contributing factor, but not a universally verifiable explanation for differing outputs. ![]()
Gentoo is where dreams and compile times go to die.
And you'll definitely find me talking to a wall :-)
Offline
Getting a repeatable answer out of a remote LLM is an exercise in futility in addition to everything that has already been mentioned you do not know if it will use settings you send such as seed, if your query is reaching the same host through a router, or for that matter when using lots of cheap routers what model it is you are being served. The host can also switch settings such as quantization, effort for peak hour throttling or more minor things such as system prompt, the host can inject what ever it wants such a the guard model flagging seth's love of smurfs and requiring the main model to verify all smurf related discussions follow the hosts TOS.
Basic facts such as "What is the capital of China" should almost always give the same output for the same model but even asking for hello world type programs can lead to minor variations even when the client is sending an identical query same seed temp 0 new session applying as much control of the other settings the remote allows without generating a 4XX response such as writing the program to disk using the shell and heredoc in one attempt, direct file system write in the next, inbuilt tool in the third, the 4th it decides to rewrite it after reading it back etc.
Last edited by loqs (Today 02:22:27)
Offline
with my reply I wanted to point out: if you ask 10 different humans you may get 10 different answers - but if you ask the same human 10 times you likely get the same answer 10 times - no matter how wrong it is (or you get an answer just once and 9 references to it - which is what i expect from a machine - although, now as i write this ... seth? by any chance you see blurry white 6502 assembler on red background - and have a casulty counter down to one decimal?)
Offline
As botnets have enough resources (IP address ranges CPU and GPU cycles) to bypass protection mechanisms such as Anubis, OTP, real email address verification and return real time results on protected content I suspect it will have no impact on scrapers wanting to index reddit.
I assume the plan is to flag users for bot-like behavior, suspend their account and dynamically quota their IP history for generating new ones, other users remain unaffected since they authorize w/ their private creds, not an IP graylist or anubis so the account requirement acts as whitelist against aggressive bot repellents.
Wait for governments to seize on that to issue digital and totally anopseudonymous human slips that you'll require to use the internet…
Offline
I assume the plan is to flag users for bot-like behavior, suspend their account and dynamically quota their IP history for generating new ones, other users remain unaffected since they authorize w/ their private creds, not an IP graylist or anubis so the account requirement acts as whitelist against aggressive bot repellents.
This assumes there is no overlap between between botnet ranges and legitimate user ranges so ISP's performing carrier grade NAT only need one botnet device and requird all users to not use VPNs.
Even if such separation can be achieved for reddit it does not work for Arch unless the project is willing to gate all project content behind a login check and even then it would make the SSO registration bottleneck worse as then the botnets through throwaway email accounts are in the queue.
Last edited by loqs (Today 19:00:23)
Offline
IT WAS BY NO MEANS MY INTENTION TO SUGGEST ARCH (or anyone) FOLLOWS THAT APPROACH!
I hope that was loud enough ![]()
I was originally merely pointing out the status qua ante against the OPs assertion and then reddits stated response to the AnnoIance and then speculating on how they might envision this to shield them from the AI-bots after you objected to their theory.
The IP ranges are actually not relevant for use, everyone w/ an active account gets in, regardless of segment.
It does indeed become an obstacle on new account registration and yes, the initial validation will become the bottleneck - also sounds familiar…
Offline
Pages: 1