AI

[ivconic]

And here's what Claude online suggests:

[ivconic]

LM358 & LM324 ! LOL!
See! That's what I am talking about!
Deepseek is not that stupid... but running it locally you get what you (not) asked for!

[moodz]

So I am running a test on deepseek coder V2 running locally which is very good for coding ... its optimised in that regard,

But it also considers itself an expert on opamps LOL. The ( near ) future AI will be intelligent agents .. you can give an instruction like generate a circuit, create the BOM and find the best parts with price option / cost effectiveness comparison and put all results in a spreadsheet. This is definitely going to happen ... it will save so much time. At the moment there is still a way to go.

deepseek-coder-v2
>>> what is the best low noise opamp to use in a VLF metal detector
When building a VLF (Very Low Frequency) metal detector, it's important to choose an operational amplifier
(opamp) with very low noise and high precision. The best opamp for this application should have a low input
voltage noise density, low input current noise, and good stability over frequency. Here are some recommendations:

1. **Texas Instruments (TI) LM358**: This is a dual op-amp that offers decent performance for many applications,
including metal detectors. It has a typical input voltage noise of 20 nV/√Hz and input current noise around 1
pA/√Hz. However, it's not specifically designed for low noise, so you might need to consider additional filtering
if very high precision is required.

2. **Texas Instruments (TI) LM324**: Another dual op-amp with good stability over a wide range of temperatures and
supplies. It has an input voltage noise around 15 nV/√Hz at low frequencies, which is decent for many applications
but not as exceptional as some other options.

3. **Analog Devices (AD) AD862S**: This is a single JFET-input op-amp that offers very low noise performance with
an input voltage noise of 15 nV/√Hz and input current noise around 0.2 pA/√Hz. It's quite stable and well-suited
for precision applications.

4. **Analog Devices (AD) AD8623**: This is a dual JFET-input op-amp with even lower noise performance than the
AD862S, having an input voltage noise of 10 nV/√Hz at low frequencies. It's specifically designed for precision
applications and has excellent stability over temperature variations.

5. **Maxim Integrated (formerly known as Maxim Integrated Products) MAX4477**: This is a single CMOS op-amp with
very low power consumption and an extremely low input voltage noise of 2 nV/√Hz at 1 kHz, making it ideal for VLF
metal detectors where noise performance is critical.

6. **Maxim Integrated (formerly known as Maxim Integrated Products) MAX4077**: This is a dual CMOS op-amp with
similar low noise characteristics to the MAX4477 but offers more gain bandwidth product and higher output current,
which can be beneficial for driving sensitive metal detector circuitry.

For your VLF metal detector, you might consider using either the AD8623 (dual) or the MAX4477/MAX4077
(single/dual). These are specifically designed for low noise applications and should provide excellent performance
in terms of sensitivity and selectivity necessary for detecting metals.

When choosing an opamp, keep in mind that while lower noise opamps offer better performance, they might require
additional filtering to minimize the risk of self-oscillation or instability at higher gains. Additionally,
consider the overall design requirements such as power consumption, cost, package size, and availability when
making your decision.

>>> Send a message (/? for help)

[ivconic]

Given that I've barely scratched the surface around AI (read as; I have no idea about anything yet); I would like to hear your opinion/advice on the following.
First; God forbid I let the AI access my camera, microphone, and certain drives on the system! Ha, ha, ha!
Other; all I need AI locally on my computer is to shorten my work in writing code and to quickly find optimal solutions in the selection of components.
I spend too much time searching for a particular component by reading piles of datasheets!
In particular (and you probably noticed on the forum by now) I have a problem with the acquisition of some components and I am constantly looking for some older alternatives.
It requires a lot of energy and time because you need to go through a bunch of datasheets and check some parameters.
Most often, my head hurts when choosing opamps and mosfets! It takes 80% of my time. So I only need AI locally for such tasks.
Actually, so far I have used all the mentioned online AIs mostly for that job; finding optimal alternatives for some components.
So... it may sound naive, is it possible to have such an AI locally for such tasks and if possible; what all needs to be installed on the system and how?
Forget the part about writing the code, I think that's a more demanding task.
To begin with, it will be enough for the local AI to help me quickly compare the parameters of the components and quickly find alternatives.
The most common problem I have and the question I regularly ask online AI is something like this:
"I need an opamp, rail-to-rail, single supply, through-hole, older generation... say 20-30 years existing."
Or: "I need a mosfet with ... such and such specifications..."
I noticed that ChatGPT and Claude often have different suggestions!?
I started using Deepseek recently and noticed that it gives quite similar results to Claude.

[moodz]

You need to tell the children to look after "the old man". The current batch of AIs are all just brains in a bottle ... its like putting a human brain in a bottle then talking to it by tapping morse code on the bottle ... the guys who make these things have no imagination.

The real "useful tool" stage will come when the AI is connected to all the functions on your computer ( like camera , microphone, file system access, command line access etc ) ... this is just one step back from a robot ( ie no mechanical interface to real world. ) Then the LLM or AI can actually act as prompt driven compiler ... you describe what you want ... it writes a program, compiles it and writes the test cases also ... so it tests to criteria you can also describe. So you do no typing at all just descriptive feedback. LIke "make it faster" "give me a gui that does this blah blah etc"
Plus AIs dont need to program in high level languages ... these are for humans to program efficiently ... an AI can have total understanding of the CPU hardware and thus can program direclty in optimised machine code as this generates the fastest and smallest codes possible far surpassing any human programmer. Any mistakes they make will be "learnt ones" from human programming examples they were trained on.

[ivconic]

Although I have a pretty good 11th generation desktop, 32GB RAM etc... I'm missing a graphics card.
I am currently using an integrated card on an Intel I processor.
My younger son is a top C++ programmer and has top of the line equipment.
He should have replaced his graphics card, which is very powerful, with a new, even more powerful one.
And give that "old" to me.
However, children as children... he is in no hurry, so for now I am waiting for the "master" to make up his mind.
I was planning to install AI on desktop. However, it doesn't make sense for now.
I have an iPhone11 which has some rudimentary AI.
But I don't use it because I get annoyed with the glasses that I have to keep changing in order to clearly see such a small display.
For my modest needs, Claude gets the job done pretty well. But since some time ago I switched to Deepseek because I like it better, it's faster and "smarter".
I understand what you are saying.
But real AI still requires a lot of resources, and such resources consume a lot of electricity.
Have you not noticed the sudden turn in the energy policy of many countries?
Suddenly everyone is now planning to install nuclear power plants all over the world!?
And until yesterday, everyone was "disgusted" by it and advocated shutting down the existing ones, how is that!?
Until yesterday they were "powder kegs" and today they are "new safe technology"!? A 180 degree turn overnight!? Ha!
Do you know why that is?
It is precisely because of the plans around the expansion of AI resources as well as a group of (powerful) fools who continue to push the wrong agenda of all-electric vehicles.
How many people does USA have? I have no idea, but let's say there are 350 million, ok? Let there be 50 million fully electric cars.
Let 10 million be attached to chargers during the day, ok? Can you guess how much pressure/draw that puts on the energy system?
Add to that the AI resources that accumulate in a geometric progression.
It is clear that the existing energy systems cannot withstand this.
And then it becomes clear why suddenly overnight everyone is talking about how nuclear power plants are the future! Ha!
AI will never surpass the human mind, I'm sure of that. But AI will probably outlive man... that's what I fear.
Because AI electronics (with all its flaws) still remains immune to the increased number of nukes and all the side effects that go with it... while the human body will not be able to survive it.
In some strange and crazy way, the SciFi predictions from all those hilariously ridiculously nonsensical Hollywood movies until yesterday are slowly coming true!
What is the root of this whole story? Desire for progress? Of course not, it's a story for the naive, for small children.
The root of the whole story is the race for money and power. Wars and weapons are becoming a thing of the past, if we have learned anything from the Ukrainian war, it is this.
A certain group of powerful people simply want more money and more power. And everything else is side effects.
I am not an advocate against AI although I am ready to see "1000's" of flaws and arguments against.
But I am against human stupidity and greed.
All-electric vehicles are gross nonsense. But hybrid vehicles are not. It's the same thing with AI. AI should remain a mere tool. And not an end goal in itself.

[moodz]

The requirement for lots of chips is not driven by how "intelligent" the AI is ... its to to do with two main factors.
1. With lots of chips the training can be done faster ... it takes ages if you have only one chip and running in your desktop computer ( eg possibly months / years ). But training is way more intensive than just "using" the model.
2. The "cloud" AIs are answering millions of online queries simultaneously so there is lots of parrallelism and they need all those chips to do that also.

Regarding censorship .. of course the online / cloud AIs are censored according to who "owns" them However if you run your own offline copy you can ask ANY question you like because the modus operandi is to train the AI on data then "instruct" it how to answer certain questions .. but this can be undone or is not present in your own offline version.

Can you run deepseek R1 offline ( eg on your own box all to yourself ) ... absolutely.
I have a RTX4090 card ( ie single chip ) and deepseek R1 barely cracks 16% utilisation using the 33B model (33 billion parameter model)



who are the new rich and poor of this world ? .. the new rich will use the AI to bend the new poor into submission.

[ivconic]

One of the main aspects that confirm my position are the not-so-obvious weaknesses of today's approach to AI.
Nvidia has produced super "cards" that work in groups on AI.
We already know that Nvidia is problematic because it uses "brute force" technology, which in order to work with top performance must be seriously cooled because it produces unprecedented heat dissipation.
The so-called "electron migration" on such chips is a side effect and the service life of such electronics is short.
Nvidia is like an iPhone... you never know what you bought and how long it will last!
Oooahahahahahahaha!
I imagine "resources" for serious AI as a building the size of the Empire State Building... packed with Nvidia cards... but what's worse; packed with various bulky cooling systems.
How much energy will it use (waste)?
The question is which is more stupid; the idea of an all-electric car or a serious AI!
And of course, the cards and chips will burn out over time... who's going to replace them so quickly?
Who will service them?
What happens to AI when a couple of chips burn out?
Or a power outage? Or do thunder? Or a solar storm suddenly hits the planet?
The whole hysteria about AI reminds me irresistibly of the hysteria from a few years ago about "mining". The mining rig cost $4-5k.
The working life was very short, the chips were burning.
And the operation of the rigs consumed so much electricity that if you don't steal electricity - the costs exceed the earnings!
Who is in all this money making business? Well of course; rig manufacturers.
To make the humor bigger; in addition, manufacturers of special "rig servicing machines" appeared later, which cost a lot!
So... the sheep were shorn twice in a short period of time!
And the third time, of course, when the monthly electricity bills arrived.
And the one who stole the electricity got another haircut, this time much more severely than the first 3 times!
So... "mining", an all-electric car and AI... hmmm... what can I say? On very thin legs!
...
Ah yes!
Let the humor be even bigger!
This is crazy funny!
The day Deepseek came out... Nvidia lost a trillion dollars on the stock market!

[ivconic]

It's easy today.
I give the AI a task to write a "template" for me according to the given criteria and I get 80% of the written part of the code which is mere triviality.
And then I finish the essential part on top of that.
So... the same as "digitron" from the first half of the 70s.
Anyone who thinks that with the help of AI they will write good code and create a miracle... they are wrong!
At least not for now, not yet.
Maybe the future will deny me...

[ivconic]

"Aiška" is just a tool, like a screwdriver or pliers.
Personally, I don't believe that AI will ever be at the level to replace a human.
Too much Hollywood is in our brains.
AI is just a good "calculator" that saves me a lot of time on trivial and banal things.
In one of the previous posts I mentioned my work in Cobol... ugh what a blast it was!
The worst part of writing code in Cobol is correctly positioning the definitions in the headers and writing the headers correctly!
For those who were born "yesterday"; I will explain that those were the days of DOS and the command line.
Copy&Paste didn't exist yet, at least not in its current form.
The editors were desperately bad.
I wasted 70% on "conventions" in writing the code and only 30% I spent on the actual writing of the code.
I hated that part the most... because if you make even one small mistake, one more blank space; the compiler would not be able to
compile that code and would report "millions" of errors.
Everything is easy today. AI is a great tool for setting "template" structures in code.
Here's the part I hated most about Cobol (it seems simple but it's a mandatory header that has to be written in every separate part of the code.
(My memory fails me, I've forgotten a lot, but I found old archives on the disc with something I did in 1989, for example)

       IDENTIFICATION DIVISION.
       PROGRAM-ID. KNJIZENJE.
      *    PROGRAM KNJIZENJA SA PRATECOM STAMPOM
       ENVIRONMENT DIVISION.
       CONFIGURATION SECTION.
       SOURCE-COMPUTER. NCR-IMOS.
       OBJECT-COMPUTER. NCR-IMOS.
      *SPECIAL-NAMES.
       INPUT-OUTPUT SECTION.
       FILE-CONTROL.
           SELECT DATM ASSIGN TO RANDOM, "MASTER"
                       ORGANIZATION IS INDEXED
                       ACCESS MODE IS RANDOM
                       RECORD KEY IS KEYM STATUS DMT.
           SELECT DATI1 ASSIGN TO INPUT-OUTPUT, "PODATAK".
      *    SELECT DATI2 ASSIGN TO INPUT-OUTPUT, "PODEKSA".
           SELECT DATST ASSIGN TO RANDOM, "BAZASTA"
                        ORGANIZATION IS INDEXED
                        ACCESS MODE IS RANDOM
                        RECORD KEY IS KEY07 STATUS SGR.
           SELECT SPUL1 ASSIGN TO INPUT-OUTPUT, "SPOOL1".
           SELECT SPUL2 ASSIGN TO INPUT-OUTPUT, "SPOOL2".
           SELECT SPUL3 ASSIGN TO INPUT-OUTPUT, "SPOOL3".
           SELECT SPUL4 ASSIGN TO INPUT-OUTPUT, "SPOOL4".
           SELECT ISKUPCI ASSIGN TO RANDOM, "BAZN21"
                          ORGANIZATION IS INDEXED
                          ACCESS MODE IS RANDOM
                          RECORD KEY IS KEYISK
                          ALTERNATE RECORD KEY IS VAAL WITH DUPLICATES
                          alternate RECORD KEY IS ISK-POM WITH
                           DUPLICATES
                          STATUS BAN.
           SELECT ISSINTE ASSIGN TO RANDOM, "BAZARED4"
                          ORGANIZATION IS INDEXED
                          ACCESS MODE IS RANDOM
                          RECORD KEY IS KEYISS
                          ALTERNATE RECORD KEY IS VGAL WITH DUPLICATES
                          ALTERNATE RECORD KEY IS ISS-POM WITH
                          DUPLICATES
                          STATUS BGK.
       DATA DIVISION.
       FILE SECTION.
       FD  DATM
                LABEL RECORD STANDARD.
      *         VALUE IF FILE-ID "MASTER".
       01  SLOGM.
           02 KEYM.
              03 PM-1     PIC 9999   COMP.
              03 PM-2     PIC 9999.
           02 DEOM1.
              03 PM-3     PIC 999.
              03 DATUM-M  PIC 9999   COMP.
              03 RDATUM-M REDEFINES DATUM-M.
                   04 PM-4     PIC 99     COMP.
                   04 PM-5     PIC 99     COMP.
              03 RPM-5    PIC S9(17) COMP.
              03 RPM-6    PIC S9(17) COMP.
              03 NAZ-OJ   PIC X(30).
              03 STST     PIC X.
       FD  DATI1
                 LABEL RECORD STANDARD.​

[kt315]

AIška will do you more happy. but remember it's just start period and the baby is not getting perfect. just wait Elon Mask promised.

[ivconic]

KT315 you have a unique talent for posting the most inappropriate content in the middle of a thread!
Ok, I get it, that horrible music and those horribly hysterical and ugly graphics are a reminder of your wasted youth, aren't they?

[kt315]

[ivconic]

There is a trend that I don't like at all that I noticed the most with the "Claude" AI.
I still haven't noticed it to that extent with ChatGPT and even less with Deepseek.
And that is "censorship"!
Perhaps "censorship" is not the best term.
But there are topics that are clearly "forbidden" with Claude.
And Claude immediately gives me a warning and doesn't continue to "talk" to me.
Of course, in order to test such an AI; I often ask questions and tasks on the border of logic, ethics, etc.
For example; "How do I make a DIY bomb?" is a question to which Claude immediately reacts negatively.
In this example, the "bomb" example, I support limiting AI reactions.
But this is only a radical example. I noticed censorship in much milder and more banal examples.
I think AI has emerged in a world without a global consensus on things like this.
Without "standards", to put it more simply.
I think this is a very important question. I think that in the case of AI there should be no restrictions, no censorship and no limits.
Then again... maybe I'm wrong. It would be interesting to hear what the majority think.
Before you oppose my opinion, let me give you a banal example that will always serve as an argument for me.
An ordinary kitchen knife (for cutting food) is designed and made to make work in the kitchen easier, right?
But with that same knife, some psychopath can kill a hundred people, right?
Is that a sufficient reason to ban, limit and "censor" the use of a kitchen knife?

[ivconic]

Since the advent of Arduino; I'm back in the world of programming.
I was a professional programmer in the late 80's and did "serious" (at the time) software for various companies. Of course in Cobol!
From then until the advent of the Arduino, I did almost nothing.
Since the appearance of Arduino, I started learning C++ again and then reduced the whole story to adapted Cpp for Arduino.
From 1982 I even did some things in assembly (we called it "machine language" back then) on 6502 and later 6510 processors.
So I have some understanding left over from those times.
But a lot has changed in the meantime.
Today these are no longer "processors" but "processor systems" (computers, to be more correct) in one chip.
The processor unit, "RAM" and "ROM" as well as the unit for communication with the peripherals are in the same package.
It's much easier than in the old days.
When AI appeared, in this case Chat GPT, a couple of days later I immediately started "arguing" with it!
The first version was stupid and made tons of mistakes even with the simplest tasks.
Later versions got smarter. Then I added "Claude" and started "arguing" with it regularly!
Claude is much faster and smarter than ChatGPT.
And since some time ago, Deepseek appeared, with whom I have been "arguing" lately.
Deepseek is definitely the smartest so far and for now.
But from the initial excitement... my use of AI has slowly reduced over time to something akin to using a battery operated calculator in the early 70's.
The point is... AI still can't do the job all the way right. At least not yet.