AI

[ivconic]

AI – Artificial Idiot

[ivconic]

Finally!
Some smarter people than me have probably managed to do this long before, I just managed to set things up right yesterday.
Offline AI on your computer, which works pretty solid!
My desktop would have been a dream configuration until yesterday, today it is already a "mediocre" machine.
I3, 10th gen, 11th gen motherboard, 32GB RAM, ATI Radeon RX 6500 XT (good but only 4GB)...ultra fast drives etc...
The power of your configuration will directly condition the AI model you can use.
In my case it is Mistral 7B Instruct v0.3.

[ivconic]

The main step is to set the AI to use GPU power. In my case it is also the main limiting factor.
What to say about this model? Works surprisingly well and fast!?? I didn't expect it!
Of course, there are flaws. The model is not able to supplement its own knowledge and store it in the database.
But there is a "temporary" solution.
For example, if I want to do more delicate work with some processor, for example Atmega328P; it's enough to give the model a datasheet and that's where the "magic" starts!
The "magic" only lasts as long as that session is active, that's the main problem with this model, because it doesn't "remember".
I noticed that the GPU resources are quite relieved even though this is not a powerful graphics card (by today's terms).
That gives me the idea to try a more powerful model.
I was too "scared" by the claims on the internet and that's why I chose this model for my configuration.
However, I see now that it does not waste too many resources.

[ivconic]

Here is an example of the main flaw of offline AI, instead of finding accurate data; AI "guess" based on the words in question:

[ivconic]

[ivconic]

To direct this AI model to the details of the Geotech forum; I would have to spend hours in the additional information I would have to give it.
Ok, that would all make sense and be very productive; if the model had the ability to remember everything and expand its knowledge.
To create its own local knowledge base.
But unfortunately, this model cannot do that.
All the "knowledge" it gathers; lasts only for the duration of the session.
Could it be that I don't know how to set it up properly?
More experienced help would be welcome!

[ivconic]

Just spent hours trying to expand this AI to be "self learning", because it is possible.
But!
Whoever invented and pushed "Python" in this sad world; IS ABSOLUE UNDISPUTED CRETIN!
CRETIN WITH ALL CAPITALS!
I even think that sad creature is much bigger CRETIN than the CRETIN who inveted Linux (Linuss idiot&cretin).
...
I will make this up to work 100% sure.
This post is just my small contribution to the ultimatet TRUTH!



P.S.
Beacuse... only a total CRETIN can figure out and make something that is NOT legacy and compatible in all recent versions.
And only an incurably ill CRETIN wants to use CMD with kilometer-long and totally meaningless instructions even today in the 21st century.
So much for that.
If someone is offended... I'M GLAD, I'M ENJOYING IT!

[moodz]

Just spent hours trying to expand this AI to be "self learning", because it is possible.
But!
Whoever invented and pushed "Python" in this sad world; IS ABSOLUE UNDISPUTED CRETIN!
CRETIN WITH ALL CAPITALS!
I even think that sad creature is much bigger CRETIN than the CRETIN who inveted Linux (Linuss idiot&cretin).
...
I will make this up to work 100% sure.
This post is just my small contribution to the ultimatet TRUTH!



P.S.
Beacuse... only a total CRETIN can figure out and make something that is NOT legacy and compatible in all recent versions.
And only an incurably ill CRETIN wants to use CMD with kilometer-long and totally meaningless instructions even today in the 21st century.
So much for that.
If someone is offended... I'M GLAD, I'M ENJOYING IT!

LOL ... you would rather it was written in C or even worse ... CPP ?

You cant really train a realistic model from scratch in a reasonable time .. without a pile of NVIDIA GPUs with heaps of VRAM. However you can fine tune them .. this is like a training a doctor to be a cardiologist.
Taking an existing model eg qwen_coder and then doing the specialist training.

There are plenty of web frameworks out there to hold your model of choice ( eg oobabooga ). Not sure why you you having trouble because you just load the framework, download the model and then put it in learning mode. Its all mouse clicks there is no python programming. ( unless you are writing your own framework ).


Theres plenty of youtubes and howtos out there ...

[ivconic]

The model works, I have no problems.
But it has no possibility of permanently accumulating what it has learned.
Everything it learned lasts only within the session.
I checked yesterday, Nvidia has support but it's hard to find for my AMD Radeon RX 6500 XT.
The newer version of LM does not have "Local Docs", instead it has a small "clipper" (symbol for attach, upload...) with which "knowledge" (PDF,DOC etc) is added to the session.
One solution is to add RAG to LM.
I looked on the net how to do it. And then the trouble starts.
Python - The biggest problem, because every version is not directly inheritable and compatible.
GIT. Visual Studio Build Tools.
RAG models... which one, suggsted was all-MiniLM-L6-v2.
Tried and failed. Every time stuck to Python incompatibility.
It all revolves around huge support for Nvidia and almost no support for AMD Radeon RX 6500 XT.
My intention was to use GPU resources and not CPU.

[ivconic]

Things are changing rapidly in the AI "arena".
Immediately after appearing in the public, DeepSeek beat ChatGPT and Claude to the ground, in all aspects.
I was thrilled for the first couple of days.
Today, a few months later, the situation is drastically more dire.
ChatGPT and Bard have improved tremendously (competition works wonders) and are now superior to DeepSeek.
These days I use AI to solve some really boring jobs.
And out of habit I turned to DeepSeek first.
I wasted several hours on the piles of nonsense and errors that DeepSeek generated.
Fortunately, I know programming and find errors and illogicalities very easily.
And then I set the same tasks to Bard... not expecting miracles.
And a miracle happened, Bard immediately gave perfect solutions to the tasks that DeepSeek was embarrassed by!
Claude is considered the best AI when it comes to programming in general.
But I didn't need to "talk" to Cloude about these tasks because Bard did it perfectly.
I'm writing some communication protocols for the Windows platform, nothing to do with detectors... so I won't go into the details here.
I just wanted to point out the drastic changes in the AI rankings. The race will speed up and become very interesting!

[ivconic]

I have "written" a dozen very good codes for various metal detectors so far.
Based on Atmega328P, ESP32... I could literally "pollute" this forum with new projects.
But that would be a poor lie, pretending to be an expert when in fact I am not...etc.
I think that such an approach to new ideas is pointless.
Second, I am a pessimist. No matter how good the code is; MD technology has exhausted all its capabilities.
The bottleneck is coil and soil as a very difficult and problematic medium to master.
So I don't want to bother here on the forum with a mediocre repetition of what the AI generated on my behalf.
I leave that to the younger enthusiasts.
I think it is much more important to academically discuss new ideas, new and original approaches.
And the mere creation of hardware and writing code, these are technical things that have become very trivial today.

[ivconic]

[ivconic]

This is not a joke.
This is literally the picture I see when I turn on the TV and watch press conferences.
I probably have some veil over my eyes, which allows me to see what is real and filters out the outer beautified facade.

[ivconic]

[ivconic]

The AI balloon is slowly deflating: