Induction Balance Stuff - Single/Multi Frequency Response, GB, Disc, Measurements, Ideas, Fun, etc.

[Aziz]

Hi all,

my new FKP2 capacitors didn't arrive. It will take some time.
But I'll rewrite the LCR meter code and finish it. And I will write mode code until my capacitors arrive. Or make a new mono coil.
Aziz

[Aziz]

Hi all,

after rewieving my old LCR meter code, I'll let the AI to code for me this time.
AI is capable to do it in full complex math with I/Q decoded vectors (measured complex voltages Vref, Vdut).
The AI gave me really very important tips too.
I also want a 4-step calibration feature (Short reference, DUT open, DUT short, DUT load - another reference load) to get maximum accuracy this time.
I want to measure the real loss of the FKP2 capacitors (BTW, not arrived yet).

I have to code the GUI for the changes yet.

[Aziz]

Regarding the very important tips from AI:
- Complex division:
I have changed the function by a numerically stable complex division according to Smith (1962) against overflow/underflow. This is absolutely required for the new LCR meter implementation.

- Phase relation may get lost by averaging filters (like EMA):
Vref, Vdut are measured complex voltages from line input. I walked into a trap by my latest changes on LCR meter implementation and was wondering about the measurement errors.
Don't do any Vref and Vdut complex averaging before calculating results. The phase relation gets lost or gets changed by the averaging filter. And hence the calculation of LCR measurement.
Filter the calculated result only (Zdut and hence L,C,D,Q,Rs,.. derived from impedance Zdut).

Does this has an important impact to the ground balance procedure too (projection method)?
Still needs to be checked.

Aziz

[Aziz]

Hi all,

regarding complex LCR-Meter with AI help:
I have tested DeepSeek AI first time. It is generating a nice C-Code. But full of bugs. I am not happy yet.
Claude Code is down at the moment. And it isn't finishing the job due to limit arrive.
Google AI is still making lots of bug. But you can ask for fixing. Ask for finding bugs. Ask for to be very very sure for correct implementation. Ask again. And again. (Up to dozen times).
I have to built-in the code and test it.

[Aziz]

Good news!
AI (Gemini and DeepSeek) failed totally to make me happy.
Wrong calculations, wrong calibrations, wrong results, ...
Even if I ask the other AI to fix the bugs.
For such a simple and trivial task.
Complex numbers does not mean it is a very very complex task and calculation.
I can't really believe it.

A.I. = Artificial Idiocracy

[boilcoil]

Hey Aziz,
Why are you surprised? AI (LLM) is a statistical processing model. It cannot calculate, it cannot think. It cannot self-learn, it cannot change itself, it cannot expand its capabilities, it cannot use the global network in real time, etc. It can only generate the best option in the transformation of the input to the output. The only thing it can do (this is the basis of its training before its release,when a model is still open, before it is "frozen" ) is - "Find which is the next token and continue the logical chain".
AI is a great tool. It is no coincidence (I think it was Moodz) that it was compared to the discovery of the wheel. However, in order to be able to use it, you must take into account all these specifics. I think that in this regard Moodz is the best.
The worst thing, however, is that LLM is not updated at all, it is simply replaced with the new version of the model, which is trained in a better way. However, upon replacement, nothing else is inherited except your personal preferences stored in your "personal long-term memory" - so the work done with LLM_V1 will not be transferred to LLM_V2 and will probably lost.

Therefore, use AI as much as possible, don't get angry with it, and never trust it without reservation.

[Aziz]

Hi all,

Claude AI isn't obviously down. It isn't working in my Firebox browser anymore. Who has made the bug?

I can't really believe it! Maybe the AI has fixed the code.
"Fix the code" Claude!

I'm so fkn fkn fkn hell disappointed.

[Aziz]

Hi all,

my new set of WIMA FKP2 capacitors came yesterday.
I do have now a lot of low capacitance values from 33 pF to 1nF. I can test the mono coil VLF with extreme high Q LC tank soon.

My PeakTech LCR meter isn't able to measure the D and Q of the FKP2 capacitors. D is showing 0.000 and Q above 2000 won't be displayed even.
Some values have Q up to 1000 or lower.
I hope, my Sound card LCR meter will give me more reasonable values for D and Q.
Aziz

[Aziz]

Hi all,

regarding LCR Meter implementation with AI help:
I have installed the Google Chrome browser. Claude Code is running there.
Claude Code is still making bugs but it can fix it.
"Fix the code" Claude!

There is a new artificial intelligent (AI) artist in the world:
"Claude Money":

It want's me to upgrade to the pro version. Or I have to wait several hours before I can go on.
Anyway, "Claude Money" makes me more happy.


Cheers,
Aziz

[boilcoil]

Hi all,
.................................................. ...........
I'm so fkn fkn fkn hell disappointed.

Hey Aziz,
I don't want to spam your topic, but if you want fewer errors and bugs - use Python. It is the most well-known language in AL. If you are interested in why LLMs make errors - read the attached article. It is useful mainly because it gives you guidelines on how to proceed most usefully with AL and what to expect from it.

[Taktyk]

Hey Aziz,
I don't want to spam your topic, but if you want fewer errors and bugs - use Python. It is the most well-known language in AL. If you are interested in why LLMs make errors - read the attached article. It is useful mainly because it gives you guidelines on how to proceed most usefully with AL and what to expect from it.

When you ask an AI for code in a popular language, the model has thousands of examples to rely on. However, when you ask for something in IBM APL2 (a niche language) or EES (a specialized DSL), the AI tries to 'guess' the solution by applying patterns from popular languages for instance, inserting FOR loops or IF-THEN-ELSE blocks where they simply don't belong. This is a classic example of hallucination based on statistics rather than a deep understanding of syntax.

In the case of languages like EES, where the logic is based on systems of equations rather than sequential assignments, the model gets lost because its training, which is based on sequential code, 'doesn't know' how to think in parallel. This demonstrates that LLM architecture has inherent limitations that cannot be overcome simply by adding more data.

If you are not an expert, you won't be able to detect that the AI has provided 'useless but plausible-looking' code. In niche programming, AI can become a 'machine for creating technical debt,' generating code that might compile, but doesn't solve the problem as intended.

Perhaps the future does not lie in large, universal models (like GPT-5 or Claude), but in smaller, specialized models (Small Language Models – SLM) fine-tuned on precise technical documentation and constrained by specific syntax (i.e., constrained generation).

As for metal detectors, this is an extremely narrow field; there are very few projects on GitHub. My experience shows that while LLMs can statistically relate to general concepts, they make a multitude of fundamental errors - for example, confusing TX frequency with ground filter frequency in Hz.

While it is possible to find your way in general DSP-related tasks, knowledge regarding Metal Detection (MD) processing is virtually non-existent in these models. My work with these models over the last month has exposed everything. If I didn't have years of experience, knowledge from patents, and insights from relevant posts on the Geotech forum, using an LLM to search for information would have been completely pointless.

[boilcoil]

Thanks, Taktyk.

[Aziz]

Hi boilcoil,

Hey Aziz,
I don't want to spam your topic, but if you want fewer errors and bugs - use Python. It is the most well-known language in AL. If you are interested in why LLMs make errors - read the attached article. It is useful mainly because it gives you guidelines on how to proceed most usefully with AL and what to expect from it.

thanks for the interesting paper.
But I don't use Python in my projects. Only C/C++. C++ option part is even kept very very low (Windows GUI related only).
DSP related stuff only in strict ANSI-C (C89 or C99). Strict procedural programming. No object oriented stuff.
So the LLM's should have a very large learing base with ansi-c code.

Claude Code is really nice. But I am facing with token limitations and penalty waiting periods (I'm using the free version only).
"Fix the code" prompt didn't get any results today. But it consumed all the free AI tokens. Next try in 5 hours.
I have to break down the task prompts into simple portions.
Aziz

[boilcoil]

.................................................. .......
I have to break down the task prompts into simple portions.

It's really a good idea to divide the task into separate parts.

.................................................. .......
But it consumed all the free AI tokens. Next try in 5 hours.

Hmm, this is really annoying.
I don't know if it would help, if you tried to work "in parallel", i.e. register another profile, necessarily with another IP (another computer, proxy...). The article above hints at this when they trying with another IP. You should have a "double resource" in such a mode of operation. The only problem would be if in the process of work up to now you gave the AI some specific data or information that it used. This "external update" will be lost in your new profile and you will have to conduct the training again.

[Aziz]

Hi all,

this is the current AI LCR meter implementation with bug fixing corrections or little optimizations.
I know, there are still minor bugs in there.
I get same results with my old implementation using calibration procedure up to Dut short (no Dut load applied).
I still can not trust in the Dut load calibaration/correction part. This part has to be fully checked yet.

Do not get confused about the LCRMETER2 structure or LCRMeter2_.. function names. I am running the old LCR meter with new LCR meter (2) at the same time to compare results.
But it shows the nice implementation using complex math with complex measured voltages Vref and Vdut straight out of the I/Q demodulator outputs.
Aziz