Request for Comprehensive RouterOS v7 Manual with Examples to build code generation chatbot

honeyfairy · Thu Jun 27, 2024 11:59 pm

I am currently working on a project that involves extensive configuration and management of MikroTik devices running RouterOS v7. I want to build a gpt/llama3 ai code generation chatbot so I need a complete manual that includes detailed command references, examples, and particularly, BGP filter configurations just for ros7

Currently the chatgpt chatbot (even the custom ones) gives me incorrect code that does not work

So I'm looking for:
A full manual or detailed guide just for for RouterOS v7 in pdf format
Command references with practical examples, especially for BGP filters for ros7

The bot will help network administrators quickly generate accurate and effective configurations, saving time and reducing errors. I can also make it publically available.

Amm0 · Fri Jun 28, 2024 2:00 am

Count me as doubtful that "accurate and effective configurations" be as simple as plugging in more specific training data, but IDK. Mikrotik does have PDF linked on the main docs page: https://help.mikrotik.com/docs/. And maybe there is a PDF for the "old wiki" someplace too, but that have old syntax too... I suppose there are training materials for certifications, but AFAIK those aren't public. But that's kinda it for docs.

It the "examples" that are more difficult to come by. I think most LLMs relay on Reddit, YouTube, and GitHub for any RouterOS knowledge. But I'd imagine this forum actually be better, or at least larger data set of RouterOS config/etc. Unfortunately there is not "export all forum posts" here to do that.

Maybe if Mikrotik published this forum's posts in some archive format (i.e. the 1,000,000 posts here, which has to more than reddit), there be more training data available.

honeyfairy · Fri Jun 28, 2024 3:26 am

Count me as doubtful that "accurate and effective configurations" be as simple as plugging in more specific training data, but IDK.

Agreed, but it should be better than whats available currently. I get terribly wrong code atm.
maybe it should crawl this forum since the time ros7 was released..

Fri Jun 28, 2024 4:37 am

Why anyone would use a Chatbot to generate routing policy is beyond me.

This is something you _need_ to understand, both from the impact that it can create on the public internet, and for the ability to troubleshoot it later.

If you don't understand what you are doing, you will never be able to troubleshoot it.

Fri Jun 28, 2024 10:40 am

It would be interesting to see if training a model on the documentation and forum would work.
here is the PDF, it's on the documentation page: https://box.mikrotik.com/f/7d18a1be34c241b4adaa/?dl=1

anav · Fri Jun 28, 2024 2:35 pm

Blasphemy, putting all consultants out of work and no more MUMs and no more youtube casters ...........................
I would be more impressed by an AI RoS Bot Tester, that would provide better and more stable code

))))

honeyfairy · Thu Jul 11, 2024 3:05 am

It would be interesting to see if training a model on the documentation and forum would work.
here is the PDF, it's on the documentation page: https://box.mikrotik.com/f/7d18a1be34c241b4adaa/?dl=1

Thank you normis, I will use this and provide feedback.

3zzy · Tue Mar 04, 2025 10:24 am

@honeyfairy were you able to use the documentationto make something useful?

I quickly tried creating a command list and a cheat sheet but it of course requires review and updates.

RouterOS v7 Cheat Sheet: https://gist.github.com/3zzy/61e356f0bf ... 6e30d80698
RouterOS v7 Commands: https://gist.github.com/3zzy/2f53dac174 ... e6b5919a67

Tue Mar 04, 2025 11:10 am

Please do not throw in the AI generated crap asking to review that "job".

3zzy · Tue Mar 04, 2025 11:16 am

Please do not throw in the AI generated crap asking to review that "job".

Not sure what you have against AI but whats been compiled could geniunely be helpful to someone new to Mikrotik and networking in general like me. AI is clearly useful in certain things and it has certainly helped me so just sharing what I could. If you don't like it, feel free to ignore?

What I shared is not a "job" and I don't get anything from your review, the community benefits.
I don't see many Cheat Sheets and Commands listed floating around. If AI has done half the job, even if its "crap" we could improve it further and make it useful.

rextended · Tue Mar 04, 2025 1:24 pm

The problem with posting s–t generated here on the forum is that, in addition to the s–t generated by real people,
it adds to the material with which intelligence is trained and users (the few) who search for something on the forum also find the remains of s–t from artificial intelligence.

Not to mention that people, that freely help, often find out later that the configuration or script for which they donated free time,
was actually made initially by artificial intelligence and must correct its mistakes,
instead of having those who use it learn something...

There is a risk that the forum, from a place of comparison and education between people,
becomes a place to correct the errors of artificial intelligence...

anav · Tue Mar 04, 2025 1:55 pm

Please do not throw in the AI generated crap asking to review that "job".
Not sure what you have against AI but whats been compiled could geniunely be helpful to someone new to Mikrotik and networking in general like me. AI is clearly useful in certain things and it has certainly helped me so just sharing what I could. If you don't like it, feel free to ignore?

What I shared is not a "job" and I don't get anything from your review, the community benefits.
I don't see many Cheat Sheets and Commands listed floating around. If AI has done half the job, even if its "crap" we could improve it further and make it useful.

I disagree, all you will do is increase the illiteracy of MT users.

3zzy · Tue Mar 04, 2025 2:30 pm

The problem with posting s–t generated here on the forum is that, in addition to the s–t generated by real people,
it adds to the material with which intelligence is trained and users (the few) who search for something on the forum also find the remains of s–t from artificial intelligence.

Not to mention that people, that freely help, often find out later that the configuration or script for which they donated free time,
was actually made initially by artificial intelligence and must correct its mistakes,
instead of having those who use it learn something...

There is a risk that the forum, from a place of comparison and education between people,
becomes a place to correct the errors of artificial intelligence...

You describe AI outputs as "s--t" multiple times but if it were truly worthless, it wouldn't pose any threat to the community. What I think is actually bothering you (and it's a legitimate concern) is that AI has become good enough to:

Create content that many users can't distinguish from human expertise
Produce solutions that work well enough to be implemented
Capture value from your knowledge and corrections without compensation

It's not that AI content is worthless, it's that your expertise, which took years to develop, is being approximated without proper recognition or compensation. And when you freely offer corrections, those improvements get fed back into making AI even better. That's a legitimate concern but doesn't justify dismissing the technology outright.

3zzy · Tue Mar 04, 2025 2:32 pm

Not sure what you have against AI but whats been compiled could geniunely be helpful to someone new to Mikrotik and networking in general like me. AI is clearly useful in certain things and it has certainly helped me so just sharing what I could. If you don't like it, feel free to ignore?

What I shared is not a "job" and I don't get anything from your review, the community benefits.
I don't see many Cheat Sheets and Commands listed floating around. If AI has done half the job, even if its "crap" we could improve it further and make it useful.
I disagree, all you will do is increase the illiteracy of MT users.

Really? How so? The AI has essentially just summarized information from the official documentation. Making concise information more accessible doesn't create illiteracy - if anything, it helps people get started who might otherwise be overwhelmed by dense technical documentation. Many people learn better with simplified explanations before diving into more complex details.

rextended · Tue Mar 04, 2025 2:40 pm

Two posts ago, the answer seemed to me to be written by AI,
completely written in an unnatural and artificial way.

And as I have written several times, I only write what I write, nothing else.
So, if your intelligence has not yet understood it, I am against posting those incorrect -hi- results.
Yes, I said incorrect. Not correct, wrong, fallacious, etc.
Do you understand the meaning?
Those who read the forum as beginners could take that -hi- for true. That's the point.
So instead of helping the beginner, it confuses him from the beginning.

rextended · Tue Mar 04, 2025 2:51 pm

[*]Create content that many users can't distinguish from human expertise

Exactly, many users, not all users...

[*]Produce solutions that work well enough to be implemented

Maybe, never seen (well enough).

[*]Capture value from your knowledge and corrections without compensation

I am among the many who have written a ton of scripts in the forum,
yet if the AI is trained with the data on this forum, I do not see the same reasoning in the AI, and not even better...

doesn't justify dismissing the technology outright.

I would also like to explain that I am not against AI, I am against those who do not understand it,
and put incomplete or incorrect results in the forum, which should only contain fairly correct answers,
not invented commands or illogical reasoning, as the chat bot often does.
The job of AI currently is to chat and invent just for the sake of chatting.

Luckily, it's not my job to provide support in RouterOS, so if there's an AI that can improve security and scripting, welcome.
You, instead, be careful not to be replaced...
You may find yourself needing to become a laborer without the ability to express yourself,
who must follow the AI's instructions to the letter, treated only as a expendable machine to do manual labor.

(And then there's no need for AI to be treated that way...)

Josephny · Tue Mar 04, 2025 3:10 pm

I very much don't want to inflame, or feed the existing flames on the topic of AI, but, in case it is of value to anyone, for the past week or so I have been working with claude.ai (their paid 3.7 version) and it is, IMHO, fantastic.

Claude.ai explains topics truly beautifully, is infinitely patient with my questions, and generates fully commented and (I believe) good quality code.

With the help of others here (you know who you are), claude.ai, a little test lab, and herculean efforts (to overcome my substantial learning disorder

), I am progressing nicely in my learning.

I would be happy to share some of the config, but only on request, as I am sensitive to the admonitions against posting AI generated code.

rextended · Tue Mar 04, 2025 3:15 pm

But look, I would call it a heartfelt discussion, rather than a flamed one.
So far it is at a high level of democracy and civilization.

3zzy · Tue Mar 04, 2025 3:19 pm

I very much don't want to inflame, or feed the existing flames on the topic of AI, but, in case it is of value to anyone, for the past week or so I have been working with claude.ai (their paid 3.7 version) and it is, IMHO, fantastic.

Claude.ai explains topics truly beautifully, is infinitely patient with my questions, and generates fully commented and (I believe) good quality code.

With the help of others here (you know who you are), claude.ai, a little test lab, and herculean efforts (to overcome my substantial learning disorder ), I am progressing nicely in my learning.

I would be happy to share some of the config, but only on request, as I am sensitive to the admonitions against posting AI generated code.

Ah It's refreshing to hear from someone who's actually explored what good AI can do! Claude 3.7 is indeed remarkable.
This discussion reminds me of the transition from paper maps to GPS navigation. When GPS first appeared, many experienced navigators dismissed it as unreliable or claimed it would make people lose their natural sense of direction. They weren't entirely wrong about some limitations, but they missed the bigger picture.
GPS helps millions navigate today. Experienced travelers still bring their expertise to interpret the GPS suggestions, understanding when to follow them and when human judgment should override. But they don't refuse to use the tool entirely.
Ultimatelt AI is still a tool - exactly as effective as its user knows how to make it.

rextended · Tue Mar 04, 2025 3:30 pm

Experienced travelers still bring their expertise to interpret the GPS suggestions, understanding when to follow them and when human judgment should override.

How many people I found going the wrong way,
children run over by those who end up on the sidewalk to look at the screen...
But it would have all happened anyway with the advent of the smartphone.

When I go for a walk now I inevitably have the GPS in my smartphone, in new cities I use it, but I haven't lost my sense of direction.
When I go by car I always have a real Garmin GPS with me, with all the personalized maps in addition to the street ones.
A truly useful tool, it doesn't lie, it doesn't invent, the information either has it, or it doesn't have it, or at most, if it does,
it was given the wrong information, but no one intelligently invented it...

Tue Mar 04, 2025 3:42 pm

As it happened to be I used CoPilot today to get some user info with specific parameters out of an Active Directory environment.
But it had some flaws which I was easily able to correct, because I know my basics.

Funny you mention GPS ... there are parts in the world where GPS doesn't work. You wouldn't believe it but it's true.
I would think France is not considered a 3th world country, however there are plenty of blind spots there in the woods.

What if you do not have a normal map with you ?
I know plenty of people really getting LOST without GPS. They don't know about the sun, where it is more or less at a given time of day. They have no sense of directions. They never learned to read a map either (not even when it's on their smartphone offline).

The basics.

Tue Mar 04, 2025 3:49 pm

With the help of others here (you know who you are), claude.ai, a little test lab, and herculean efforts (to overcome my substantial learning disorder ), I am progressing nicely in my learning.

Which some time ago I asked you about (and again, it was and still is a genuine honest question) and you said there was no such problem ?

Larsa · Tue Mar 04, 2025 3:55 pm

Yeah, AI is great for boilerplate work, but you need to know the basics to fix the flaws. Plus, you need access to multiple LLM engines (ie different models) since answers still vary a lot depending on the situation.

Amm0 · Tue Mar 04, 2025 3:56 pm

If you're messing with Claude, it might be worth it to try "The Dude" chatbot at https://mikrotik.com/support/. It can answer network questions like "what a trunk port?", and it was also trained on Mikrotik's docs. While I don't think pre se be "better" that just using Claude/etc, but it's starting context may be better (i.e. it know you're likely going to ask a networking or Mikroitk question).

But it's RouterOS scripting where these LLM falter. Stuff like JavaScript is 10000x more common & on top there is plenty of linter to pre-check output of LLM for JavaScript validity. I'd have to imagine more LLM take advantage of "LSP" (see https://langserver.org), while open standard for presenting code in editor it can be consulted at any point in AI workflows. But since there isn't a LSP, all the LLM have to "fly blind" on what's actually valid RouerOS script/config.

So if you ask me... better AI with RouterOS config start with an LSP. And LSP be useful since a real person can use it check their own code and get "hints" etc. Win, win.

3zzy · Tue Mar 04, 2025 4:00 pm

As it happened to be I used CoPilot today to get some user info with specific parameters out of an Active Directory environment.
But it had some flaws which I was easily able to correct, because I know my basics.

Funny you mention GPS ... there are parts in the world where GPS doesn't work. You wouldn't believe it but it's true.
I would think France is not considered a 3th world country, however there are plenty of blind spots there in the woods.

What if you do not have a normal map with you ?
I know plenty of people really getting LOST without GPS. They don't know about the sun, where it is more or less at a given time of day. They have no sense of directions. They never learned to read a map either (not even when it's on their smartphone offline).

The basics.

I understand the point you're making about knowing the basics, and I agree that foundational knowledge is valuable. However, I think you're focusing on edge cases to dismiss tools that benefit the vast majority of users in the vast majority of situations.
Yes, GPS has blind spots in some remote areas. And yes, having backup skills is sensible. But we don't structure our entire approach to navigation around these exceptions. Most people navigate successfully with GPS 99% of the time, even if they can't triangulate their position using the sun's angle and a stick in the ground.
This is like arguing we shouldn't use cars because sometimes they break down, or we shouldn't use calculators because batteries die. Of course we should still teach basic math and walking! But dismissing tools that dramatically improve efficiency for most people in most situations because of rare edge cases isn't practical.
AI tools aren't perfect - I never claimed they were. But they're already good enough to provide substantial benefits to many users, while continuing to improve rapidly. The reasonable approach is to use them where they help while maintaining critical thinking, not to reject them outright because they occasionally make mistakes or have limitations.
Pointing out that tools sometimes fail isn't an insight - it's stating the obvious. The more interesting question is how to best integrate these tools into our learning and working processes.

Larsa · Tue Mar 04, 2025 4:04 pm

But it's RouterOS scripting where these LLM falter. Stuff like JavaScript is 10000x more common & on top there is plenty of linter to pre-check output of LLM for JavaScript validity. I'd have to imagine more LLM take advantage of "LSP" (see https://langserver.org), while open standard for presenting code in editor it can be consulted at any point in AI workflows. But since there isn't a LSP, all the LLM have to "fly blind" on what's actually valid RouerOS script/config.

Yeah, you can also clearly see the difference in answer quality based on how much training data is available. Things like Cisco and Juniper get really solid answers, probably because of their extensive documentation and all the forums discussing them.

Tue Mar 04, 2025 4:05 pm

What a discussion !!!

3zzy · Tue Mar 04, 2025 4:16 pm

What a discussion !!!

Well, you started it!

Josephny · Tue Mar 04, 2025 4:27 pm

With the help of others here (you know who you are), claude.ai, a little test lab, and herculean efforts (to overcome my substantial learning disorder ), I am progressing nicely in my learning.
Which some time ago I asked you about (and again, it was and still is a genuine honest question) and you said there was no such problem ?

The smiley emoji indicates it was a joke. And yes, it referenced your comments.

Josephny · Tue Mar 04, 2025 4:52 pm

If you're messing with Claude, it might be worth it to try "The Dude" chatbot at https://mikrotik.com/support/. It can answer network questions like "what a trunk port?", and it was also trained on Mikrotik's docs. While I don't think pre se be "better" that just using Claude/etc, but it's starting context may be better (i.e. it know you're likely going to ask a networking or Mikroitk question).

But it's RouterOS scripting where these LLM falter. Stuff like JavaScript is 10000x more common & on top there is plenty of linter to pre-check output of LLM for JavaScript validity. I'd have to imagine more LLM take advantage of "LSP" (see https://langserver.org), while open standard for presenting code in editor it can be consulted at any point in AI workflows. But since there isn't a LSP, all the LLM have to "fly blind" on what's actually valid RouerOS script/config.

So if you ask me... better AI with RouterOS config start with an LSP. And LSP be useful since a real person can use it check their own code and get "hints" etc. Win, win.

Interesting idea to try the dude chat.

I am in no way an expert at ros but I would really like to hear an expert’s analysis of Claude’s configuration advice.

It sure seems to me that it is getting the config correct, and I’ve been asking it for more and more complex solutions.

3zzy · Tue Mar 04, 2025 4:56 pm

I am in no way an expert at ros but I would really like to hear an expert’s analysis of Claude’s configuration advice.
It sure seems to me that it is getting the config correct, and I’ve been asking it for more and more complex solutions.

+1. I'd like to here an experts analysis of this which is from Claude after incorporating @CGGXANNX's suggestions: viewtopic.php?p=1130715#p1130715

jaclaz · Tue Mar 04, 2025 5:12 pm

Only to show how old I am, something very similar to this AI thingy was once called an "expert system".
You feed the system with correct data, and then it can make every kind of cheatsheets, digests and provide correct answers to correct questions.
The approach works just fine.

Most of these new AI's have been instead fed with anything that was available, shaken and mixed well, and then the expected result is correct answers.
It is a logical impossibility.

The answers may be correct (or they may be not) but not only there is no way (without checking directly) to know, but there isn't even an approximate estimation of the probabilities, instead everything is provided with a somehow "absolute" tone, the AI - probably in good faith - thinks it is right, but it doesn't really know.

Very likely - before or later - it will evolve (actually better programmed) to the point of being capable of understanding when it is talking bu**sh*t and shut up or becoming capable to simply admit it is not so sure about something, but till then its only meaningful use is for anything that is:
a. not important (like funny or not funny pictures)
b. not otherwise "exact" (like programming or scripting)

Not that humans are largely particularly good at programming or scripting, let alone document what they do.

Josephny · Tue Mar 04, 2025 7:25 pm

Only to show how old I am, something very similar to this AI thingy was once called an "expert system".
You feed the system with correct data, and then it can make every kind of cheatsheets, digests and provide correct answers to correct questions.
The approach works just fine.

Most of these new AI's have been instead fed with anything that was available, shaken and mixed well, and then the expected result is correct answers.
It is a logical impossibility.

The answers may be correct (or they may be not) but not only there is no way (without checking directly) to know, but there isn't even an approximate estimation of the probabilities, instead everything is provided with a somehow "absolute" tone, the AI - probably in good faith - thinks it is right, but it doesn't really know.

Very likely - before or later - it will evolve (actually better programmed) to the point of being capable of understanding when it is talking bu**sh*t and shut up or becoming capable to simply admit it is not so sure about something, but till then its only meaningful use is for anything that is:
a. not important (like funny or not funny pictures)
b. not otherwise "exact" (like programming or scripting)

Not that humans are largely particularly good at programming or scripting, let alone document what they do.

Not "old," but rather "wise and experienced."

I am actually not a big fan of AI at all, and am blown away at the fiction it creates in the context of providing fact.

I sum it up as AI simply being in its infancy.

But, like everything, it is a far more complex situation and I think one such complexity is that AI even at this stage is pretty good a some things (while horrible at others).

I remain very happily surprised at how well it understands and explains VLANs and ROS.

Larsa · Tue Mar 04, 2025 7:42 pm

So if you ask me... better AI with RouterOS config start with an LSP. And LSP be useful since a real person can use it check their own code and get "hints" etc. Win, win.

Since today's LLMs also understand syntax definitions and can be trained accordingly, it would be great if Mikrotik could publish an official BNF for RouterOS scripting for this purpose. If Mikrotik uses Yacc/Bison or a similar tool internally, it might not be too much work to provide a BNF.

If MikroTik published an official BNF for RouterOS scripting, it would help developers and tools, including LLMs, better understand and work with the scripting language. This would lead to smoother implementations of automation tools for network management and monitoring, enable better code analysis in external IDEs, and provide more useful learning resources for the MikroTik community.

Ps..
I plan to open a thread and make it an official feature fequest.

Josephny · Tue Mar 04, 2025 7:48 pm

So if you ask me... better AI with RouterOS config start with an LSP. And LSP be useful since a real person can use it check their own code and get "hints" etc. Win, win.

Since today's LLMs also understand syntax definitions and can be trained accordingly, it would be great if Mikrotik could publish an official BNF for RouterOS scripting for this purpose. If Mikrotik uses Yacc/Bison or a similar tool internally, it might not be too much work to provide a BNF.

If MikroTik published an official BNF for RouterOS scripting, it would help developers and tools, including LLMs, better understand and work with the scripting language. This would lead to smoother implementations of automation tools for network management and monitoring, enable better code analysis in external IDEs, and provide more useful learning resources for the MikroTik community.

Ps..
I plan to open a thread and make it an official feature fequest.

You both clearly know far far more than I do about AI, LSP, LLM, ROS, etc.

I have an opinion nonetheless (they're free, after all): Why not check out claude.ai's current level of understanding before concluding how to make AI more accurate and useful for ROS config help?

My analysis on how well claude.ai does with ROS is not worth the electrons used to express them here, but you both, and many others here, could form a more much valuable analysis of the current state -- and thereby a better path forward.

anav · Tue Mar 04, 2025 7:57 pm

I remain very happily surprised at how well it understands and explains VLANs and ROS.

No better than the references already provided by others, I think you are more comfortable/receptive with a sexy AI presence ;-P )

Tue Mar 04, 2025 8:01 pm

As friendly discussion evolves I want to mention that quotting whole provious post just to add a short comment would need soon an AI

to extract crucial information what the topic is about. Some use good old readable themes and as the topic starts to be longer and longer it gets harder and harder to scroll down dozens of screens to follow posts. There are separate "Post reply" and "Quote" buttons.

Larsa · Tue Mar 04, 2025 8:38 pm

I have an opinion nonetheless (they're free, after all): Why not check out claude.ai's current level of understanding before concluding how to make AI more accurate and useful for ROS config help? My analysis on how well claude.ai does with ROS is not worth the electrons used to express them here, but you both, and many others here, could form a more much valuable analysis of the current state -- and thereby a better path forward.

Atm, we use the pro versions of Copilot, Claude, Codestral, and Grok, but none are really accurate at ROS scripting (yet). They're great for boilerplates and documentation, though.

Josephny · Tue Mar 04, 2025 8:45 pm

The scripting inaccuracy is not surprising.

Have you tried it for ROS configuration?

anav · Tue Mar 04, 2025 8:48 pm

configuration=script

Josephny · Tue Mar 04, 2025 9:00 pm

configuration=script

I was using the same definition of script as used here when we have separate forum for scripting.

Not entirely the same as configuration.

jaclaz · Tue Mar 04, 2025 9:09 pm

configuration=script

Not really, configuration = tiny subset of scripting language (without variables, conditional execution, etc.).

It even misses (it would be IMHO very useful) the equivalent of a REM statement, or - if you prefer - a way to keep # lines in a section/directory and make item/line comments more evident (in CLI) .

Even Monicelli

https://github.com/esseks/monicelli
has the bituma statement besides #.

anav · Tue Mar 04, 2025 9:15 pm

True dat config is a very small subset of scripting , Syntax ( the word I should of used) of scripting is huge and is limited in its use on the config.

rextended · Tue Mar 04, 2025 9:20 pm

Maremma cinghiala..............

phascogale · Wed Mar 05, 2025 12:18 am

Create content that many users can't distinguish from human expertise

This statement is two-edged.

To me it is also the critical junction, like waves and tides on sands. You can see ocean on one side, dry land the other, so it is our worry, capacity and virtue to analyse the uncertainty between.

Also, @jaclaz beat me to it on expert systems. The debates today are similar for the current descendants of that software technology. I am reminded also of how everything is now "AI" in the same way it was "e-" then "i-"

Incidentally, the quick AI translation did a terrible job of dealing with @rextended's "Maremma cinghiala", even wondering whether it could be Indonesian rather than Italian.

Wed Mar 05, 2025 12:52 am

Even Monicelli
https://github.com/esseks/monicelli
has the bituma statement besides #.

Santa Vergine, Dio del cielo ... COBOLLA reinventati !!!

anav · Wed Mar 05, 2025 2:54 am

Incidentally, the quick AI translation did a terrible job of dealing with @rextended's "Maremma cinghiala", even wondering whether it could be Indonesian rather than Italian.

No worries, it was gibberish, I didnt understand the meaning of the Italian translation either '=)

jaclaz · Wed Mar 05, 2025 3:07 am

No worries, it was gibberish, I didnt understand the meaning of the Italian translation either '=)

That's ok, also many Italians not born in Tuscany would have troubles understanding It, they would of course understand the words, but not fully appreciate the meanings (that sentence can have more than one depending on context and tone).

rextended · Wed Mar 05, 2025 10:44 am

If I tried to translate it into English it would be "Come on, I don't believe it!"

anav · Wed Mar 05, 2025 2:38 pm

Let me provide some context.............

In response to "AI will soon replace rextended, the baddass script guru"
If I tried to translate it into English it would be "Come on, I don't believe it!"

..

The Tuscan Phrase revisited: Maremma cinghiala..............

anav · Wed Mar 05, 2025 2:41 pm

No? Okay howbout.......

Let me provide some context.............

In response to "I met a woman at a bar last night, nudge nudge wink wink, the next morning she told me her name was Giorgi M"
If I tried to translate it into English it would be "Come on, I don't believe it!"

....
The Tuscan Phrase revisited: Maremma cinghiala..............

Wed Mar 05, 2025 2:44 pm

Boring ... yawn ...

anav · Wed Mar 05, 2025 3:37 pm

Luv those that complain but dont offer anything better............. booo!

jaclaz · Wed Mar 05, 2025 3:54 pm

If I tried to translate it into English it would be "Come on, I don't believe it!"

Yep, but it all depends on the tones, and how they can be rendered with punctuation, that would probably be:
Maremma cinghiala ?
your:
Maremma cinghiala ...
sounded to me more like "o tempora o mores" or in English something like "Shame on this age and on its lost principles!"
and of course, not to be confused with:
Maremma cinghiala!
that could be translated to "Damn it!"

With the addition of a simple but before it:
Ma maremma cinghiala!
it becomes much stronger showing exasperation.

anav · Wed Mar 05, 2025 4:11 pm

Never wouldve guessed you are a historical cunning linguist.

Wed Mar 05, 2025 4:53 pm

Luv those that complain but dont offer anything better............. booo!

Left or right? Canadians first

anav · Wed Mar 05, 2025 9:15 pm

Dont laugh, my spouse says Im turning into them!!!

dang21000 · Wed Mar 05, 2025 9:52 pm

I agree, help is a very good evolution of the old wiki.

Work is done, and must be continued.
Wireless is at the beginning, too complicated because there OLD and NEW system (change since 7.13).
Example : searching capsman, we found info in both parts...

I'think it will better to don't upgrade the documentation like always fresh doc.... it's easier to have a doc per version, like ansible do, or netapp.
With this "per version" doc, no more problem and this help customer to find the right infos.
This is a good way to have clear doc, without thing saying "since v...." like https://help.mikrotik.com/docs/spaces/R ... OSpackages

And it must stay "commons" docs, like upgrade paths from X to Y, changes, etc...
Because upgrading wireless to wifi 7.12 > 7.13 and newer, wasn't easy...
Find, reagarding the changelogs of ros versions, how to change our conf to enable it... like new fasttrack6 with 7.18, at this time except forum, or experimanted people, who can easily find what actions do to enable it ?

Or how to know before upgrading from 7.15 to 7.18, why before i can change cpuspeed, or something else... and now i can't due to "fu...ng" device mode evolution : https://help.mikrotik.com/docs/spaces/R ... evice-mode
There no way to understand that in this doc.

I'm working with lot of san/nas devices... everytime i upgrade, i know what to do, deprecated functions, newly features, what will changes, etc... i don't discover anything after.

rextended · Thu Mar 06, 2025 2:03 pm

MikroTik has the habit of leaving the same name for different things... the user-manager of v6 has nothing to do with the one of v7, like CAPsMAN, etc...

anav · Thu Mar 06, 2025 4:04 pm

The MT AI bot should be able to spit one out ............... given time.

3zzy · Sat Mar 08, 2025 9:33 am

Probably something like this would be useful: https://rlama.dev

levicki · Sat Mar 08, 2025 4:14 pm

You describe AI outputs as "s--t" multiple times but if it were truly worthless, it wouldn't pose any threat to the community.

You got that backwards -- if it wasn't sh*t then it wouldn't pose a treat but would be genuinely useful.

- Create content that many users can't distinguish from human expertise

No, more like create content that many users can't verify at the speed and scale that is prohibitively expensive even for actual domain experts to verify. In this case it is also placing a burden on existing MT professionals that now have to waste time verifying AI answers and dealing with people complaining that AI "solutions" don't work.

Produce solutions that work well enough to be implemented

In networking and especially in network security there's no such thing as "well enough" -- you either have a 100% correct setup or you are risking intrusion, becoming part of botnet, or worse, being a victim of identity theft or your resources being used for criminal activity which can even make you liable or at least greatly inconvenience you until you have proven it wasn't you.

Capture value from your knowledge and corrections without compensation

Everything the LLM models were trained on was STOLEN. Not a dime of compensation was paid to anyone, nor has anyone's consent been asked before STEALING. And I am talking here about copyright which you do not relinquish when you publish something -- this post of mine is copytighted by me despite being posted on the public message board. I am not giving permission to train ML on its contents. If you do it, you are breaking existing copyright law. The law may be bad, but that's how it currently works, and all those LLMs have broken it on a massive scale.

I know that the people like you who evangelize AI will now jump to defend this act by saying that it doesn't differ from how humans learn and let me tell you right away that's not true. We humans buy a book to read its contents or borrow it from a public library which is paying a compensation to authors to allow that to happen. LLMs like Meta's llama were trained on PIRATED books downloaded from the torrent sites.

It's not that AI content is worthless, it's that your expertise, which took years to develop, is being approximated without proper recognition or compensation. And when you freely offer corrections, those improvements get fed back into making AI even better. That's a legitimate concern but doesn't justify dismissing the technology outright.

First, the solution to "making AI even better" isn't throwing more data in it. Every process has its limits and LLMs have reached their limit in what they can do. What should justify dismissing the technology is that whatever it produces is the fruit of the poisonous tree or if you prefer more colorful analogies then money they earn by selling subscriptions to it is blood money -- ill-gotten gains and all that.

Finally, the output is as good as input -- it will never get better or otherwise alchemists would have been able to turn lead into gold long time ago.

3zzy · Sat Mar 08, 2025 5:38 pm

You describe AI outputs as "s--t" multiple times but if it were truly worthless, it wouldn't pose any threat to the community.
You got that backwards -- if it wasn't sh*t then it wouldn't pose a treat but would be genuinely useful.

- Create content that many users can't distinguish from human expertise
No, more like create content that many users can't verify at the speed and scale that is prohibitively expensive even for actual domain experts to verify. In this case it is also placing a burden on existing MT professionals that now have to waste time verifying AI answers and dealing with people complaining that AI "solutions" don't work.

Produce solutions that work well enough to be implemented
In networking and especially in network security there's no such thing as "well enough" -- you either have a 100% correct setup or you are risking intrusion, becoming part of botnet, or worse, being a victim of identity theft or your resources being used for criminal activity which can even make you liable or at least greatly inconvenience you until you have proven it wasn't you.

Capture value from your knowledge and corrections without compensation
Everything the LLM models were trained on was STOLEN. Not a dime of compensation was paid to anyone, nor has anyone's consent been asked before STEALING. And I am talking here about copyright which you do not relinquish when you publish something -- this post of mine is copytighted by me despite being posted on the public message board. I am not giving permission to train ML on its contents. If you do it, you are breaking existing copyright law. The law may be bad, but that's how it currently works, and all those LLMs have broken it on a massive scale.

I know that the people like you who evangelize AI will now jump to defend this act by saying that it doesn't differ from how humans learn and let me tell you right away that's not true. We humans buy a book to read its contents or borrow it from a public library which is paying a compensation to authors to allow that to happen. LLMs like Meta's llama were trained on PIRATED books downloaded from the torrent sites.

It's not that AI content is worthless, it's that your expertise, which took years to develop, is being approximated without proper recognition or compensation. And when you freely offer corrections, those improvements get fed back into making AI even better. That's a legitimate concern but doesn't justify dismissing the technology outright.
First, the solution to "making AI even better" isn't throwing more data in it. Every process has its limits and LLMs have reached their limit in what they can do. What should justify dismissing the technology is that whatever it produces is the fruit of the poisonous tree or if you prefer more colorful analogies then money they earn by selling subscriptions to it is blood money -- ill-gotten gains and all that.

Finally, the output is as good as input -- it will never get better or otherwise alchemists would have been able to turn lead into gold long time ago.

It still reeks of someone pissed about compensation. Read your own words again - it's all about compensation and stolen data. Yet here's what you're missing: AI is indeed good enough - or I would say damn near close to perfect in many domains.
They probably haven't even focused on networking yet, let alone specifically on MikroTik. If they did train AI thoroughly on networking protocols and MikroTik configurations, it would definitely provide near-perfect solutions to 80% of setups. Your "100% correct or nothing" argument ignores that humans make configuration errors constantly - the difference is AI can be systematically improved.
Your argument about verification is just moving goalposts - first it's "AI is shit," then it's "well, it's too hard to verify." That's like saying calculators are worthless because someone needs to verify the inputs.
The theft argument might have legal merit, but it doesn't change the technical reality. AI is getting better regardless of how you feel about its training. And honestly, if your expertise can be "approximated" this easily, maybe the real concern isn't the AI.

anav · Sat Mar 08, 2025 5:44 pm

The concerns are legit........... stealing work, creating false information, etc............
Which means it needs to be regulated 100%

Technology, science etc. can be used for good and evil purposes and if you think everybody is ethical and moral, you are not living on this planet and are blind to history.
Sure its a good thing, if developed and used wisely.
Would free up my time if 80% solutions were good, and I could focus on the 20%, not sure if people would actually learn much but I suppose the naturally curious would and the rest would continue to live their dull lives.

jaclaz · Sat Mar 08, 2025 6:03 pm

Which means it needs to be regulated 100%

Yep

, so we can tax it

:
https://www.marriedtothesea.com/040709/ ... dillos.gif

anav · Sat Mar 08, 2025 6:29 pm

I am not discussing the implementation of regulation which is a separate topic, but simply the need for it.
Same as the need for regulations and standards for medical research for example or flight safety etc.....
I also dont have much time for people that whine about things instead of providing solution paths for an implementation that makes sense.
In other words a short quip about tax, is nothing but a useless distraction.

levicki · Sat Mar 08, 2025 7:10 pm

It still reeks of someone pissed about compensation. Read your own words again - it's all about compensation and stolen data. Yet here's what you're missing: AI is indeed good enough - or I would say damn near close to perfect in many domains.

And your argument seems to imply the issues I pointed out are nothing more than sour grapes.

As for the "good enough" claim -- that depends on how much you know about those many domains which, judging by how you praise AI, doesn't seem to be much. In short, typical Dunning-Kruger in effect.

If they did train AI thoroughly on networking protocols and MikroTik configurations, it would definitely provide near-perfect solutions to 80% of setups.

Thing is, those 80% of setups can be covered by default manufacturer configuration. People who come here and use Mikrotik do not fall in those 80%.

Your "100% correct or nothing" argument ignores that humans make configuration errors constantly - the difference is AI can be systematically improved.

The difference is that a human posting a configuration example will do so based on their experience so it will be a copy of a working and tested setup. Decent human beings will also check what they share before posting, perhaps even make a quick lab setup to make sure it works. Not to mention that if you pay someone for advice like you are paying the LLM subscription they are obligated to provide a proper service.

Your argument about verification is just moving goalposts - first it's "AI is shit," then it's "well, it's too hard to verify."

It's not either / or -- it's a causative relationship. AI is sh*t because it is hard to verify even for people who have the domain knowledge, let alone for people who are clueless.

That's like saying calculators are worthless because someone needs to verify the inputs.

That's not even remotely similar. If you type 2+2 into a calculator it WILL always output the same result 4. Ask the AI the same thing several times and confront it and suddenly you will have several differently worded responses. In order to understand why that's a problem you'd need to understand how LLMs work which seems to be out of reach of your mental capabilities or you wouldn't be singing this blind praise.

The theft argument might have legal merit, but it doesn't change the technical reality. AI is getting better regardless of how you feel about its training.

So your argument is "the end justifies the means"?

And honestly, if your expertise can be "approximated" this easily, maybe the real concern isn't the AI.

My expertise is not in danger at all as what you call "AI" (i.e. a LLM) will never be able to write the code like I do -- my sanity and patience are because AI generated trash has permeated all spheres of life.

For example, some 3D asset creatores have started using AI generated images as textures for 3D assets and are selling those assets for $15. In reality, the "quality" of those generated textures makes them not worth buying, let alone paying that much. Worse yet, those assets are not labelled as AI generated and I now have to wade through a quagmire of that crap to find useful stuff.

Did you try Google search recently? Not just the AI summaries, but AI answers (most of which are flat out wrong), but the veritable torrent of AI written "articles" abusing SEO to farm ad revenue making it all but impossible to find what you are looking for on the first page or even at all.

And you call that progress? All I see is ensh*tification of everything.

3zzy · Sat Mar 08, 2025 7:25 pm

It still reeks of someone pissed about compensation. Read your own words again - it's all about compensation and stolen data. Yet here's what you're missing: AI is indeed good enough - or I would say damn near close to perfect in many domains.
And your argument seems to imply the issues I pointed out are nothing more than sour grapes.

As for the "good enough" claim -- that depends on how much you know about those many domains which, judging by how you praise AI, doesn't seem to be much. In short, typical Dunning-Kruger in effect.

If they did train AI thoroughly on networking protocols and MikroTik configurations, it would definitely provide near-perfect solutions to 80% of setups.
Thing is, those 80% of setups can be covered by default manufacturer configuration. People who come here and use Mikrotik do not fall in those 80%.

Your "100% correct or nothing" argument ignores that humans make configuration errors constantly - the difference is AI can be systematically improved.
The difference is that a human posting a configuration example will do so based on their experience so it will be a copy of a working and tested setup. Decent human beings will also check what they share before posting, perhaps even make a quick lab setup to make sure it works. Not to mention that if you pay someone for advice like you are paying the LLM subscription they are obligated to provide a proper service.

Your argument about verification is just moving goalposts - first it's "AI is shit," then it's "well, it's too hard to verify."
It's not either / or -- it's a causative relationship. AI is sh*t because it is hard to verify even for people who have the domain knowledge, let alone for people who are clueless.

That's like saying calculators are worthless because someone needs to verify the inputs.
That's not even remotely similar. If you type 2+2 into a calculator it WILL always output the same result 4. Ask the AI the same thing several times and confront it and suddenly you will have several differently worded responses. In order to understand why that's a problem you'd need to understand how LLMs work which seems to be out of reach of your mental capabilities or you wouldn't be singing this blind praise.

The theft argument might have legal merit, but it doesn't change the technical reality. AI is getting better regardless of how you feel about its training.
So your argument is "the end justifies the means"?

And honestly, if your expertise can be "approximated" this easily, maybe the real concern isn't the AI.
My expertise is not in danger at all as what you call "AI" (i.e. a LLM) will never be able to write the code like I do -- my sanity and patience are because AI generated trash has permeated all spheres of life.

For example, some 3D asset creatores have started using AI generated images as textures for 3D assets and are selling those assets for $15. In reality, the "quality" of those generated textures makes them not worth buying, let alone paying that much. Worse yet, those assets are not labelled as AI generated and I now have to wade through a quagmire of that crap to find useful stuff.

Did you try Google search recently? Not just the AI summaries, but AI answers (most of which are flat out wrong), but the veritable torrent of AI written "articles" abusing SEO to farm ad revenue making it all but impossible to find what you are looking for on the first page or even at all.

And you call that progress? All I see is ensh*tification of everything.

You're really good at putting words in my mouth while ignoring the actual technical reality. The Dunning Kruger accusation is ironic coming from someone dismissing an entire technology because it threatens their worldview.

Let's be clear about a few things:

The 80% figure isn't about "default configurations" - it's about the complex, custom setups that MikroTik users actually need. What you're missing is that pattern recognition is exactly what AI excels at, and networking configurations follow patterns.
Your "working and tested" human example is idealistic fantasy. Have you seen the garbage configurations people post online? At least AI improves systematically over time, whereas human errors persist indefinitely.
Your calculator comparison misses the point entirely. AI isn't deterministic by design - that's a feature, not a bug. It's meant to adapt to different contexts and requirements, just like human expertise does.
The "end justifies the means" strawman is deflection. The legal framework for data usage is still evolving, but that's separate from whether the technology works.
Your skills may be safe for now, but your attitude reeks of the same dismissiveness that typists had toward word processors or switchboard operators had toward automated exchanges.

So feel free to keep raging against the machine while the rest of us figure out how to use it productively.

Amm0 · Sun Mar 09, 2025 12:35 am

And why the robots are likely going to win in the end... Us humans are going kill ourselves arguing.

All I know is docs could be improved, in many ways... which help robots and humans.

rextended · Sun Mar 09, 2025 2:36 am

typists had toward word processors

switchboard operators had toward automated exchanges

Two examples that require determinism, against something not determinate, as you yourself wrote.

I don't want a word processor to correct my sentences as someone else likes, just because someone else whose LLM has stolen the work has to put it in mine...

I don't want a switchboard to call me from a different number than the one I dialed, because usually at that time I call someone and they think I've got the wrong number...

AI is unreliable in principle precisely because in the exact same context it gives different solutions.

To do 2+2, you have to do 2+2 (ignoring how it is actually done by the calculator circuit, always the same).
I don't care that today 2+2 is 16/8*32/16 and tomorrow it tells me that it is √16 leaving me to check why it didn't write directly 4.
Most of the produced codes are full of useless operations, it looks like Windows source code...
Not to mention that he never admits that he doesn't know how to do it, making things up,
if an command doesn't exist, and no document provided has it, he makes it up from scratch... just to please the writer.
They are just chat machines, and it is obvious that you are into them, you probably have no one else to talk to.

3zzy · Sun Mar 09, 2025 3:43 am

typists had toward word processors

switchboard operators had toward automated exchanges

Two examples that require determinism, against something not determinate, as you yourself wrote.

I don't want a word processor to correct my sentences as someone else likes, just because someone else whose LLM has stolen the work has to put it in mine...

I don't want a switchboard to call me from a different number than the one I dialed, because usually at that time I call someone and they think I've got the wrong number...

AI is unreliable in principle precisely because in the exact same context it gives different solutions.

To do 2+2, you have to do 2+2 (ignoring how it is actually done by the calculator circuit, always the same).
I don't care that today 2+2 is 16/8*32/16 and tomorrow it tells me that it is √16 leaving me to check why it didn't write directly 4.
Most of the produced codes are full of useless operations, it looks like Windows source code...
Not to mention that he never admits that he doesn't know how to do it, making things up,
if an command doesn't exist, and no document provided has it, he makes it up from scratch... just to please the writer.
They are just chat machines, and it is obvious that you are into them, you probably have no one else to talk to.

Your fixation on determinism misses the entire point of modern AI. Deterministic tools like calculators solve simple, well-defined problems. Complex domain knowledge isn't deterministic - even human experts disagree on optimal approaches.
The analogies weren't about determinism, they were about resistance to technological change.
Your examples are telling ... you don't want a word processor correcting your sentences? That's literally what millions of people rely on daily with spell check and grammar tools. And they work because they're probabilistic, not deterministic.
I don't talk to AI because I'm lonely, I use it as a tool to augment my capabilities, just like every other technology we've adopted.

New tech disrupts established patterns. But your hostility suggests something beyond technical concerns. AI will improve where it's useful and fade where it isn't. That's how technology evolution works. Your emotional resistance won't change that trajectory.

rextended · Sun Mar 09, 2025 4:26 am

Your emotional resistance won't change that trajectory.

How cute, the emotions, the love, the envy, the joy...
It has nothing to do with it. Do you want to explain whether or not a black hole exists.. with a fart???

You are fear of this, you are fear of that, you are envious of this, you are envious of that...
Usually one speaks as one thinks, and your fixation on the feelings of fear and envy makes me understand, without the need for AI, that your life is full of them.
You always have them in mind...

Even the obsession with compensation...
As if you were missing something and your subconscious attributes to others the need for compensation that you feel inside...

It seems that you are helped by artificial deficiency to write the answers, yourself do not even know how to read.
First I contested two of "your" examples, which are not appropriate, then I did not write that "the spell checker should not correct typing errors",
but I wrote, (and it is very specific what I wrote, it is not that every time you reread the sentence the words change),

I don't want a word processor to correct my sentences as someone else likes, just because someone else whose LLM has stolen the work has to put it in mine...

There is no need for me to explain the sentence, it is clear what is written.
If you can't understand that, get help from artificial deficiency.

I already know your little game.
Anyway, the earth is flat, and whatever you write, I can surely prove the opposite without having to write anything.

I didn't think I'd find anyone worse than the artificially deficiency at writing and writing without saying anything.
Now I understand why you like the AD...

rextended · Sun Mar 09, 2025 4:53 am

I'll try to explain better what concerns me, so that no assumptions can be made.
If I forgot something, ask...

Q) What would change in my life if an AI came out that wrote scripts better than me and others here on the forum?
A) Absolutely nothing to me.

Q) If the aforementioned AI existed, would I use it?
A) Apart from mere curiosity, I prefer to do things by myself, because I like to keep my mind trained and I don't want it to atrophy like many morons I see around (I'm not referring to here on the forum).
It takes me less time to do what I need, rather than generating something that I then have to double check, and in case of problems, since I didn't write it, I would have to go and look for where the defect is and it would take me much longer...

Q) If a generic AI that worked existed, would there be some aspect that would make me lose my job?
A) No, my job is based on human relationships, and people are already fed up with automatic voices or call centers. They want to talk to the real people in charge.
And then for the "physical" part that I do, I don't see how any audio/video/text/etc. generated by the AI could replace me (or my colleagues).

Q) If there was an AI that would make me money, would I use it?
A) Of course, I'm not an idiot...

Q) What are your concerns about the world if a working generic AI came out?
A) People are already sheep, I don't see what could possibly get worse.
At most, those people who will become useless will lose their jobs, since they are always in the office without doing anything physical, who can easily be replaced.
In the end there will only be a bunch of laborers who will have to blindly follow the instructions on a display, who will not have the ability to argue or understand what they are doing.

As a matter of fact, already now, the majority of those who use a Smartphone now do not even know why the access point, if it is wireless, still has a power supply...
Already now when people see something on the internet they think it's true... just because they saw it on the internet...

One day when we put AI everywhere, even in cars, robots, military planes, etc.,
it will be so intelligent that it will understand that the ruin of the world is humanity, and as man kills rats,
AI will start killing humans, so useless are they now that there is a supreme being who knows everything and can do everything.

phascogale · Sun Mar 09, 2025 5:10 am

... you don't want a word processor correcting your sentences? That's literally what millions of people rely on daily with spell check and grammar tools.

@3zzy, your example is perfect, in the sense that millions of people commit daily blunders of spelling and grammar simply because they accept what the software tells them; they know no better, or do not stop to consider.

This then feeds into the problems of expert time wasted and of IP theft (not meaning an address).

And they work because they're probabilistic, not deterministic.

Exactly wrong. They "probably" work meaning there is an error rate. You can count on that, it is practically determined.

Expert systems, core configurations, standard processes, analytical and generative aids all have roles The combination speciously labelled AI is a further tool, not a panacea, and is morally burdened.

rextended · Sun Mar 09, 2025 5:20 am

This thing made me think so much,
that I could write a book about it.
A story of a civilization so advanced that no one leaves home,
everyone is in a cabin and no one has ever really seen anyone other than their family and the robotic doctor.
Everyone would look the way they want to others.
People would only meet virtually.
Then towards the end the twist, in reality they were not at home, but inside a spaceship for generations, headed towards a new planet to colonize, and all the experiences lived and the people met were fake and generated by artificial intelligence.
It turns out in reality that in the spaceship only two families were left alive (coincidentally him & her) and they will have to revive humanity on that new planet...

mozerd · Sun Mar 09, 2025 11:17 am

This thing made me think so much,
.............
.............
It turns out in reality that in the spaceship only two families were left alive (coincidentally him & her) and they will have to revive humanity on that new planet...

@rextended, you are in the right TRACK

... makes for an excellent movie script ...

anav · Sun Mar 09, 2025 4:34 pm

Already made sort of --> WALL-E and of course the MATRIX, where the AI were smarter than rextended and didnt just kill us but used us for power generation LOL

Mon Mar 10, 2025 8:05 am

... you don't want a word processor correcting your sentences? That's literally what millions of people rely on daily with spell check and grammar tools.
@3zzy, your example is perfect, in the sense that millions of people commit daily blunders of spelling and grammar simply because they accept what the software tells them; they know no better, or do not stop to consider.
...

It's a perfect example showing how simple human abilities are falling down for millions of people.
Honestly, with hands on your harts, how many of you can write with a traditional pen a few pages of text with no errors, in a readable way?
Even simple signing documents becomes harder and harder.
Reading ... oh man ... what are the YT, TT and other platforms for? "Do not want to read!, I WANT TO WATCH AND LISTEN TO!"

Mon Mar 10, 2025 8:51 am

Reading ... oh man ... what are the YT, TT and other platforms for? "Do not want to read!, I WANT TO WATCH AND LISTEN TO!"

I truly and profoundly HATE youtube instruction videos.
Total waste of time in most cases.
I can read faster then they can explain it in a video. Even in English.
And when it's a digital manual, even better. Then I can search.

(with one exception: instruction videos e.g. to replace guiding wheels for a printer

That you can not describe clearly enough)

levicki · Mon Mar 10, 2025 7:19 pm

You're really good at putting words in my mouth while ignoring the actual technical reality.

I did no such thing -- I quoted what you said and responded to each specific bit unlike you.

You quoted everything I said so it looks like you are responding to me, and then you didn't address any of the points I made.

The 80% figure isn't about "default configurations" - it's about the complex, custom setups that MikroTik users actually need. What you're missing is that pattern recognition is exactly what AI excels at, and networking configurations follow patterns.

"If all you have is a hammer, everything looks like a nail." -- Abraham Maslow, 1966 (but there are variations dating back to 1868.

Your "working and tested" human example is idealistic fantasy. Have you seen the garbage configurations people post online? At least AI improves systematically over time, whereas human errors persist indefinitely.

And where, pray tell, will your "AI" read the correct information to improve itself once it ingested the incorrect one? Unless the original author updates their post and "AI" is retrained on it how do you expect it to improve? Also, what incenitive will there be for anyone to post anything useful if it will all be gobbled up by a machine without compensation? Me thinks your "AI" is an idealistic fantasy, not what I said. You fail to understand that if "AI" is trained on a sum of human knowledge it can only be as-good-as that knowledge, never better. But by all means, keep drinking the cool aid from "Open" "AI" and the likes.

Your calculator comparison misses the point entirely. AI isn't deterministic by design - that's a feature, not a bug. It's meant to adapt to different contexts and requirements, just like human expertise does.

It wasn't my calculator comparison -- it was yours. I just made it make sense, but you don't seem to like the result. Second, what you call "AI" is deterministic -- if you pass the same random seed you will get the same output every time.

The "end justifies the means" strawman is deflection. The legal framework for data usage is still evolving, but that's separate from whether the technology works.

Even if the technology works (and it definitely doesn't work the way you seem to believe hence my Dunning-Kruger remark), there's no legal framework to justify Meta pirating 82TB of books to train it.

But I see that the ethics is a concept you are unfamiliar with so it doesn't surprise me you are confusing it with words such as strawman and deflection. It's almost as if the subject of ethics wasn't present in the dataset you were trained on, must be modern western education.

Your skills may be safe for now, but your attitude reeks of the same dismissiveness that typists had toward word processors or switchboard operators had toward automated exchanges.

I am dismissive because I understand the limitations of LLM since I had to understand how it works in order to properly evaluate its usefulness (or the lack of it).

You on the other hand sound like a groupie for a handful of rich techbros who are creating your personal Torment Nexus while you're paying for the privilege.

So feel free to keep raging against the machine while the rest of us figure out how to use it productively.

Good luck with that, can't wait for you to tell us your IP address after "AI" creates your "secure" configuration.

Amm0 · Mon Mar 10, 2025 8:45 pm

Reading ... oh man ... what are the YT, TT and other platforms for? "Do not want to read!, I WANT TO WATCH AND LISTEN TO!"
I truly and profoundly HATE youtube instruction videos.
Total waste of time in most cases.
I can read faster then they can explain it in a video. Even in English.
And when it's a digital manual, even better. Then I can search.

I don't mind the YT videos per se. What's annoying is I know between scripting and editing, it a lot more work than maintain good docs and other WRITTEN communication. Just for example... take @normis video on Let's Encrypt, from two years ago. Anyone who watched that video be left with expired certificates today. So instead of update the docs with some canonical "Mikrotik approved" way/script of renewing certs, which isn't very hard (and while some ACME TLS auto-renew is ideal — docs can fill in those gaps BUT don't). Instead, we've got more videos and device-mode* (*instead of being able to use TLS out-of-the-box).

Maybe, we can focus on lacking of Mikrotik's written docs here, instead?

Like I've tried to highlight that is BOTH fixable (by Mikrotik) and helpful to us human and robots. And we can fight the AI wars another day.

phascogale · Tue Mar 11, 2025 4:46 am

In relation to my earlier post that it could be a useful tool but not a panacea, it is worth pausing over AI's learning ability.

AI systems are trained (I will not comment further on their training sources) then presented as is. Their knowledge base, basically an ability to associate stored tokens, is static. Tell it a new fact and find it is not listening. It is in stasis until entire retraining, which is expensive. Learning such as it is comprises modifying weightings based on personal preferences which are stored in a separate database, like usual marketing data. The fact that it changes its responses when you reframe queries, improving their precision, is a function of new associations from new seeds. The user, not the AI tool, is honing in on the correct answers.

Given that ROS configuration and scripting are or can be codified, training is possible and from Mikrotik's point of view may provide a quite useful tool for new users. However, its errors are baked in for a given prompt and it will not learn from competent users who might try to correct it. If a 'ROS AI' existed then it could be used as a prompt to assist learning by people. That is the current state and near future of the market.