Surprised at very poor chatgpt responses. Voltage divider question.

[carl0s]

I wanted to check something very basic, but I was not exactly 100% sure myself. I knew a voltage divider worked on the ratio of two resistors, but I wasn't sure of the formula, or if there was some reason why it might not be recommended for measuring a sensor output (say a motor driver's RPM output signal).
So I asked ChatGPT.
I think it's done a pretty terrible job.
Right away I could see that it gave a formula (which I didn't doubt was probably correct, and I did look up later on during the chat) but then it gave values that didn't match it's own formula.

My default first response to ChatGPT is often "are you sure about that?"..

Anyway, what do you think? Chat link below.
https://chat.openai.com/share/f06e978b-4338-4886-94b9-955e2a0297d1

[David Aurora]

I've been playing around with it a bit lately and have been surprised at how confidently it is wrong a lot of the time, which I guess is a pretty good impersonation of human behaviour.

It's like before asking it something you already have to know the answer, then you need to do this silly dance with it where it gives you an answer, you tell it that's the wrong answer, it apologises for the mistake and hopefully the next answer it gives is correct.

Will be interesting to see how it is a year down the track

[thm_w]

The top google results gives you good information:
https://www.seeedstudio.com/blog/2019/10/09/voltage-dividers-everything-you-need-to-know/
https://study.com/learn/lesson/voltage-divider-circuit-rule-bias-formula.html

The first link actually has all the information you are asking for.

chatgpt is for boilerplate, not finding trustworthy information or learning new things, IMO.

[vk6zgo]

It seems that it doesn't learn from its interactions with people, except within that interaction.
It then forgets all that & only relies on what it can glean by "looking on the 'Net".

Go back & ask it the same question in a week & it will probably make the same mistake, or a different one, but almost certainly, a mistake.

[brucehoult]

The really strange thing is it was MUCH better three months ago.

[ataradov]

It seems that it doesn't learn from its interactions with people, except within that interaction.

It does not, and it is a good thing. Microsoft Tay learned from its interactions, and it became racist nazi in a matter of days.

And yes, it looks like the quality of responses goes down over time. Something there is broken.

[themadhippy]

maybe dementia is slowly creeping in

[golden_labels]

As always, a reminder: ChatGPT is a language model. Not a knowledge model or information retrieval system. It does not “know” what it says. It’s a random text generator, constructed in a way making the output statistically likely to match patterns observed in text authored by humans. An amazing achievement, but one should never expect from it anything a language model is not good for. It’s not magic, it’s not a simulation of human mind.

Please understand this! Otherwise you are misleading yourself and falling into traps like this, but also missing a great opportunity of seeing the importance of GPTs.

ChatGPT uses a portion of the earlier conversation as its prompt. But as far as we know the model is static and does not directly depend on interactions with other users. One explanation is avoiding the repetition of the Microsoft Tay incident, but it’s not neccessarily the primary one. The size of the the model and its learning set makes a similar attack much harder. A more likely explanation is that the teaching process is expensive. But of course this is a proprietary model, so we may never learn the truth.

The move from 3.x line to 4.x made the model different and I feel like 4.x is worse in some aspects. But within the same model I did not observe responses getting worse. It seems that the censorship part is being constantly improved, blocking some interactions accepted before, but not the GPT model itself. If you’re talking about the change within the same model, I would first try to exclude a more plausible explanation: the initial shock diminishing and operators getting accustomed to the tool. This inevitably includes perception of things getting worse.

EDIT

For comparison this is a conversation I got, using 3.5 and different client:⁽¹⁾

Q: What is needed to read a 0 - 12v sensor with a 3.3v input ADC of a microcontroller?
A: To read a 0-12V sensor with a 3.3V input ADC of a microcontroller, you will need a voltage divider circuit. This circuit consists of two resistors, where one is connected between the sensor output and ground, and the other is connected between the ADC input and the output of the first resistor. The resistor values can be calculated based on the voltage levels and desired ADC range to ensure proper scaling and voltage compatibility.
Q: What are typical resistor values for this?
A: Typically, for a voltage divider circuit to interface a 0-12V sensor with a 3.3V input ADC, you can choose resistor values in the range of 10k ohms to 100k ohms. The specific values can be determined based on the desired voltage scaling and the input impedance of the ADC. It's advisable to consult the microcontroller's datasheet and perform calculations to ensure accurate voltage measurements.

At this point the answer is right. 10 kΩ and 100 kΩ resistors are sufficient to scale 12 V to 1.09 V, which is within the range of a 3.3 V input. So here is the continuation over a different path:

Q: This under utilizes the range of the 3.3 V ADC.
A: Apologies for not following your request. To utilize the full range of the 3.3V ADC, you can use an op-amp based amplifier circuit. This can amplify the voltage from the sensor before feeding it into the ADC. The specific resistor values will depend on the amplification factor required for your application. Calculations and component selection should prioritize the desired voltage range and accuracy.
Q: Can this be done by only using the right resistor values?
A: Yes, by selecting the appropriate resistor values in the voltage divider circuit, you can achieve proper voltage scaling and utilize the full range of the 3.3V ADC. The resistor values can be calculated based on the desired voltage range of the sensor and the input range of the ADC. It's important to ensure that the voltage across the ADC input stays within its limits to maintain accurate measurements.
Q: What should be the values of these resistor to use the full scale of the ADC?
A: To utilize the full-scale range of a 3.3V ADC with a 0-12V sensor, you can use a voltage divider circuit with resistor values of approximately 27k ohms for the upper resistor and 10k ohms for the lower resistor. This ratio will effectively scale down the input voltage to the ADC range while maintaining accuracy. Remember to consider the tolerance and power rating of the resistors for practical implementation.

This is a response matching a valid one.

Of course this does not prove ChatGPT in any way “understands” this, has any mental perception of this circuit or can see it critically. No, it’s nothing more than a glorified “Markov bot” with some additions.⁽²⁾ It did not even “notice” my underutilization prompt was wrong⁽³⁾ and made continued into generating text about an unrelated solution. The takeaway here is, that even minor variation in the prompt provides enough of the noise to push the response into a different direction.

⁽¹⁾ The client also provides contextual information in the prompt and adds some other modifications, but the effect of these were minimized in this conversation. Client-specific parts not included in quoted prompts.
⁽²⁾ Supposedly ChatGPT has access to some databases and calculators, making it possible for the results to be incorporated into responses. But even if, this would be entirely controlled by a language model, not human-like comprehension of the subject. It is also not a classic Markov bot, despite ChatGPT describing itself as one, but the concept is not far away. The scale is uncomparably different.
⁽³⁾ The initial answer from ChatGPT was already valid: it contained a range, not specific values.

[Mechatrommer]

i'm not surprised, do you? .. even cortana the artificial general intelligence designed in 27th century, service life is only 7 years before rampancy kicks in.. btw.. Define love.. if its love to one thing while willing to sacrifice or çause misery to other human.. then dog lover has a heart and brain of a dog, same to others lower being animal lovers.. so chagpt lover has a heart and brain of a chatgpt.. just saying..

[tggzzz]

I wanted to check something very basic, but I was not exactly 100% sure myself. I knew a voltage divider worked on the ratio of two resistors, but I wasn't sure of the formula, or if there was some reason why it might not be recommended for measuring a sensor output (say a motor driver's RPM output signal).
So I asked ChatGPT.
I think it's done a pretty terrible job.
Right away I could see that it gave a formula (which I didn't doubt was probably correct, and I did look up later on during the chat) but then it gave values that didn't match it's own formula.

My default first response to ChatGPT is often "are you sure about that?"..

Anyway, what do you think? Chat link below.
https://chat.openai.com/share/f06e978b-4338-4886-94b9-955e2a0297d1

You shouldn't have been surprised at that.
You should have thought about what an LLM is doing internally and why they are referred to as "statistical parrots".
You should have been reading news reports, e.g. https://www.bbc.com/news/world-us-canada-65735769 and many many other places

A New York lawyer is facing a court hearing of his own after his firm used AI tool ChatGPT for legal research.

A judge said the court was faced with an "unprecedented circumstance" after a filing was found to reference example legal cases that did not exist.
...
"Six of the submitted cases appear to be bogus judicial decisions with bogus quotes and bogus internal citations," Judge Castel wrote in an order demanding the man's legal team explain itself.

"Is varghese a real case," reads one message, referencing Varghese v. China Southern Airlines Co Ltd, one of the cases that no other lawyer could find.

ChatGPT responds that yes, it is - prompting "S" to ask: "What is your source".

After "double checking", ChatGPT responds again that the case is real and can be found on legal reference databases such as LexisNexis and Westlaw.

[magic]

Good one

If you think about it, maybe it says something about lawyers that a "statistical parrot" asked to generate a legal text would make some shit up and then repeatedly defend it. Hmm...

People get in trouble when they don't know what it really means for a language model to be "accurate"

[TimFox]

In that infamous legal case, where a lawyer thought he was using a search engine to look up precedents, the gpt returned answers that were in proper format, but totally fictional.
This would have been useful for a scriptwriter writing a Perry Mason episode.
The judge was not happy, since no one had checked to see if any of the alleged precedents existed in fact.
I was surprised, since I thought lawyers had access to legitimate specialized search engines to look for precedents and previous court decisions.

[KE5FX]

You should have thought about what an LLM is doing internally and why they are referred to as "statistical parrots".

Jeez, I hate that phrase.  Get back to us when a parrot writes a credible epilogue to a Fitzgerald novel, writes some code that compiles and runs successfully, or composes original poetry.  Dismissing "language models" is an indication that you, like the rest of us, don't really understand what language is and what it does.  We are not that special.

And the fact is, ChatGPT 4 was better when it came out... much better.  Either efforts to render it more politically correct have impaired its inference abilities in unrelated areas, or something else is going on that no one outside OpenAI (sic) is willing or able to discuss.

You should have been reading news reports, e.g. https://www.bbc.com/news/world-us-canada-65735769 and many many other places

No, you should not be reading news reports to gain an understanding of emergent capabilities in language models, any more than you should be reading news reports to come up to speed on superconductor research or any number of other popular topics that can't be distilled to a headline, lede, and two or three paragraphs of fluff. 

[abeyer]

Jeez, I hate that phrase.

Seems like a succinct, evocative, and accurate analogy. 

[tggzzz]

You should have thought about what an LLM is doing internally and why they are referred to as "statistical parrots".

Jeez, I hate that phrase.  Get back to us when a parrot writes a credible epilogue to a Fitzgerald novel, writes some code that compiles and runs successfully, or composes original poetry.  Dismissing "language models" is an indication that you, like the rest of us, don't really understand what language is and what it does.  We are not that special.

To parrot something is to repeat something without understanding. That matches LLM output. Statistical implies a random unpredictable component. That matches LLM output.

Get back to us when LLMs "understand" their output. Being able to do arithmetic would be a start. Get back to us when LLMs have done more than strings plausible words together.

OTOH I don't disagree that LLMs make as much sense as profoundly ignorant and/or stupid people.

And the fact is, ChatGPT 4 was better when it came out... much better.  Either efforts to render it more politically correct have impaired its inference abilities in unrelated areas, or something else is going on that no one outside OpenAI (sic) is willing or able to discuss.

Are you turning this towards an anti-PC diatribe?

You should have been reading news reports, e.g. https://www.bbc.com/news/world-us-canada-65735769 and many many other places

No, you should not be reading news reports to gain an understanding of emergent capabilities in language models, any more than you should be reading news reports to come up to speed on superconductor research or any number of other popular topics that can't be distilled to a headline, lede, and two or three paragraphs of fluff.

You should read reports that show how LLMs capabilities are both hyped and misunderstood.

[SiliconWizard]

Language models are just that, language models. If mastering a language (not that ChatGPT or any AI language models do really master any language this far, but they are certainly getting better and better at that) was all it took to be able to answer any question, I think we would know by now.

The interesting part here is that people's expectations, for anything called "AI" these days, seem way higher than even the most aggressive marketers would hope for.

Oh, and "OpenAI" is about as open as a rock and gets most of its funding from MS.

[KE5FX]

Yadda, yadda, yadda.  I hear a lot of empty talk but I don't see any parrots writing epilogues, code, or poetry.   

[tggzzz]

Yadda, yadda, yadda.  I hear a lot of empty talk but I don't see any parrots writing epilogues, code, or poetry.   

Ah. A troll.

Nobody has made those claims for parrots, so your point is (to be charitable) a strawman argument.

They have used a traditional concept (a person parroting something) to highlight what LLMs manage to do.
https://www.collinsdictionary.com/dictionary/english/parrot
2. verb If you disapprove of the fact that someone is just repeating what someone else has said, often without really understanding it, you can say that they are parroting it.

[KE5FX]

(Shrug) Polly just translated a bunch of Python 2 code to Python 3 for me and didn't even demand a cracker, so I'm not complaining.

[carl0s]

Yes I have found it very useful for knocking up boilerplate Python / Django, and refactoring various JavaScript functions. It's hopeless at voltage dividers though obviously 😁

[SiliconWizard]

This has been talked about again and again, that's because they've integrated much of the work done wifh Copilot (from github), and that could be done because programming languages are close enough to natural languages that common generative approaches can be taken. So for spitting programming code, it obviously works better than for anything else engineering.
Still doesn't mean that the code it generates is correct. It's very easy to make it generate bogus code as well.