Don’t ask ChatGPT about PLCC44 sockets

[etik61]

I just asked ChatGPT about the mapping between the package pins and the pins of a THT socket. In the first reply, the counting started at the corners. After I corrected it several times, the text became correct, but the ascii drawing still had the numbering starting at the corners. No way to fix it. The results for mapping to the THT were very confusing. An "intelligent" answer could have been a recommendation for an SMD socket or using a different package.

[Karel]

Don’t ask ChatGPT about PLCC44 sockets

Thanks but I already don't ask chatgp anything...

[thm_w]

Google PLCC44 socket, first result is a datasheet with the pin numbers: https://app.adam-tech.com/products/download/data_sheet/198233/plcc-44-at-data-sheet.pdf
Is that not what you want?

[exmadscientist]

ChatGPT and friends are fine if you're bored and frustrated and need to blow off a little steam. But never ask them a question you don't know the answer to already (or at least that you can't readily verify once you see it).

Because they're wrong. Mostly just wrong.

I've only learned something new once. (I wanted to know what the jacket on SEOOW cable is and after much pain the bot said "Santoprene". I don't actually know if that's right but it is extremely plausible and nigh-impossible to find out other ways so we'll go with it.) I have gotten volumes of garbage.

My favorite so far was asking the new "super-AI" Claude why there are no depletion PMOS transistors for sale. This answer is out there on the Internet but extremely hard to find and usually requires piecing together things from various sources, so it's a good torture test. It did eventually get there... kind of. It just kept repeating a bunch of semiconductor physics babble from some academic or two. It obviously didn't understand shit because that babble was relevant but could all simplify dramatically... but not by Claude's "brain". Comical to prompt it like I would a student... and just get the same reams and reams of crud back, a half-dozen times.

This sort of AI might be OK for the most basic of basic stuff. But it can never do anything novel. Not even plugging tab A into slot A, so long as it hasn't seen that before. Dreadful things.

[jpanhalt]

SEOOW:

S for Service. E for Thermoplastic Elastomer. OO for Oil-resistant insulation and Oil-resistant jacket. W for Weather and Water resistance. Rated to: 600 Volts.

A J instead of W means junior or 300V insulation.

Santoprene is a trademark:

The name Santoprene was trademarked in 1977 by Monsanto, and the trademark is now owned by Celanese.

Chat was wrong again.  A standards definition is never a trademark.

[thm_w]

Yeah, SEOOW is TPE, which can be various materials. Santoprene is TPV.
I thought the same that Sxxxx would define a specific material but it does not, only the class/performance. Manufacturers can vary the actual material type used a fair bit.

So if you asked "what material the jacket is" maybe that is confusing..

[Psi]

There's a tendency when one gets older to start ignoring everything said by younger people because they don't have the same years of wisdom that you do and they get things totally wrong.   
This is a fatal mistake in life, and it's the same with asking AI things.

You should ask AI a lot, but it should be obvious that, if you ask about something obscure that it didn't get much training on, the answer is going to be pretty terrible.
The same as it is when talking to a human with limited knowledge of a subject.
AI is not a tool, tools do one job well, AI is an entirely new thing.
The main issue is that AI can't tell when it doesn't know enough to answer a question accurately, so it's confident about everything including stuff it's wrong about

[tggzzz]

It's not a "tool that does a job", it's an entirely new thing.

It is designed to assemble plausible sequences of words.

Some people assemble similar "word salad"; the term for such people is "bullshitters".

[Psi]

It's not a "tool that does a job", it's an entirely new thing.

It is designed to assemble plausible sequences of words.

Some people assemble similar "word salad"; the term for such people is "bullshitters".

That's true, but you can get a lot of understanding baked into a model by feeding a shit ton of data into it and having it map out all the probabilities that link things and concepts to other things and concepts.
It becomes more than the sum of it's parts. (I'm not saying it 'thinks', it's still just a bunch of probabilities and is not even close to general AI). 

In many ways it's so much more useful than having a big library of all its training data that you have to go read yourself. But in other ways it's much worse.
The good thing is you get both, you have the AI and Internet and you get to pick which is better for what you are doing, but if you never pick the AI option you are missing out.

[Bud]

Does this thing ever say "I do not know" ?

[coppercone2]

chat gpt is extremely good at making inappropriate corporate art 

you ask it for a logo, it gives you... george carlin if he was a artist

if I designed those things I am assured someone might smirk and say "you know that your sick right?"

kinda like jokers helmet in full metal jacket. is it a corporate mascot or a taliban mortar crew in the Kandahar valley. i suppose business is war

[exmadscientist]

Yeah, SEOOW is TPE, which can be various materials. Santoprene is TPV.
I thought the same that Sxxxx would define a specific material but it does not, only the class/performance. Manufacturers can vary the actual material type used a fair bit.

So if you asked "what material the jacket is" maybe that is confusing..

Exactly, the standards don't define the exact polymer (or even generally what it is) and I wanted to know what is actually used in practice. (I forget if I wanted to glue to it, was generally just curious, or was sick of smelling the stuff that morning.) The manufacturers don't really bother to tell you. So I asked the robot that ingests half of the Internet every day for breakfast if it had ever come across anyone who had ever said anything and it came up with Santoprene. I have no idea if that's truly correct (because, remember, dumbass robot) but I have never found any better answer, and Santoprene is a damn good match, based on its general properties, the fact that it's in widespread manufacture, and the other samples/examples of it that I have.

So I don't blindly trust the robot. You can never blindly trust the robot. You can only ask it questions whose answers you can verify some other way.

[exmadscientist]

Does this thing ever say "I do not know" ?

They have dire problems with saying "I don't know": https://www.mindprison.cc/p/the-question-that-no-llm-can-answer

(Of course, this starts to make a bit more sense when you think of them as the super-turbo-nitrous-charged Markov chains that they really are.)

[SiliconWizard]

Why use ChatGPT when a search engine would do perfectly, while at least showing you what the source of the information is?

[fzabkar]

I use Copilot. It provides references. However, I mostly ask it the kinds of questions whose answers I could find with a search engine, except that Copilot could do the same thing for me much quicker. As long as you know and accept its limitations, an AI powered search can be useful.

[Psi]

Does this thing ever say "I do not know" ?

Only if you ask it a question about something where that thing had a lot of people saying I don't know in the training data. 

There is work being done to improve AI and it's ability to check its own facts and then respond based on it's confidence in them.
but yeah, it's a work in progress.
It's hard because there's a lot of wrong info that it gets trained on because the datasets are way too huge to human review. And even if you could human review them, humans don't agree most of the time.

Why use ChatGPT when a search engine would do perfectly, while at least showing you what the source of the information is?

- It's often hard to sift through all the info when searching or find the right search query.
AI provides an interface that improves this by an insane amount. Once you have the AI answer you have a lot more info to use for crafting your search to find more technical/factual info if you need it.

- Sometimes you are just refreshing your memory or just looking for a rough idea about something, and you're not worried about it saying something wrong because you will know instantly.

[tggzzz]

It's not a "tool that does a job", it's an entirely new thing.

It is designed to assemble plausible sequences of words.

Some people assemble similar "word salad"; the term for such people is "bullshitters".

That's true, but you can get a lot of understanding baked into a model by feeding a shit ton of data into it and having it map out all the probabilities that link things and concepts to other things and concepts.
It becomes more than the sum of it's parts. (I'm not saying it 'thinks', it's still just a bunch of probabilities and is not even close to general AI). 

In many ways it's so much more useful than having a big library of all its training data that you have to go read yourself. But in other ways it's much worse.
The good thing is you get both, you have the AI and Internet and you get to pick which is better for what you are doing, but if you never pick the AI option you are missing out.

Volume of data does not automatically make information, let alone understanding.

Asking whether a machine can think/understand is as useful as asking whether a submarine can swim.

Bullshitters sometimes get things right, sometimes wrong.

LLMs are about software imitating the suboptimal properties of wetware. Software should do the things software is better doing than wetware. And vice versa.

[Brumby]

Volume of data does not automatically make information, let alone understanding.

It's called "waffle".  It is ubiquitous.  It is almost entirely useless and simply wastes the time of the reader.

High school essays are particularly susceptible.

[SiliconWizard]

Volume of data does not automatically make information, let alone understanding.

Yes, and yet this is the whole premise of current "AI" (not saying that it covers all of "AI", which is yet to be really defined anyway, but just saying that there's a lot more that could be done than just mashing tons of data through CNNs).

But I've heard many, many people convinced that, on the contrary, there's nothing more to "intelligence" than a lot of data combined together. Which is convenient business-wise, as there is tons of data available, with still very little regulation.
And, from an astrophysicist's POV, everything is just information anyway, probably, so just "data". Probably. What's information though.

[tooki]

High school essays are particularly susceptible.

In a way, I was always envious of people that can do that. I’ve got no trouble writing copious amounts of text if there is a lot of information to put in, but I’ve never been good at padding.

[tggzzz]

..."AI", which is yet to be really defined anyway, ...

AI has had many many "definitions" over the decades. The common property is "apparently magic, to be viable next year."

Strangely Igor Aleksander's 1983 WISARD, effectively the forerunner of today's LLMs, wasn't included in the contemporary definition of AI. WISARD demonstrated a key property of modern LLMs: you didn't and couldn't predict/understand the result it would produce.

Example: when trained to distinguish between tanks and cars, it worked well in the lab but failed in the field. Eventually they worked out it was distinguishing between sunny days (car advert pictures) and cloudy days (tanks in Germany).

[Psi]

Volume of data does not automatically make information, let alone understanding.

Yes, and yet this is the whole premise of current "AI" (not saying that it covers all of "AI", which is yet to be really defined anyway, but just saying that there's a lot more that could be done than just mashing tons of data through CNNs).

But I've heard many, many people convinced that, on the contrary, there's nothing more to "intelligence" than a lot of data combined together. Which is convenient business-wise, as there is tons of data available, with still very little regulation.
And, from an astrophysicist's POV, everything is just information anyway, probably, so just "data". Probably. What's information though.

It depends on how you define words.
Sometimes people use the word "understand" in a more organic/alive sort of way. While others use it in more of a general logic sense.
eg, if you create a software app that has all the formulas and input/output boxes for doing your taxes does that app now understand tax because the programmer coded understanding into it?  Some people will argue yes, others no.
A lot of disagreements on the internet come down to people having slightly different internal definitions of words

[tggzzz]

Volume of data does not automatically make information, let alone understanding.

Yes, and yet this is the whole premise of current "AI" (not saying that it covers all of "AI", which is yet to be really defined anyway, but just saying that there's a lot more that could be done than just mashing tons of data through CNNs).

But I've heard many, many people convinced that, on the contrary, there's nothing more to "intelligence" than a lot of data combined together. Which is convenient business-wise, as there is tons of data available, with still very little regulation.
And, from an astrophysicist's POV, everything is just information anyway, probably, so just "data". Probably. What's information though.

It depends on how you define words.
Sometimes people use the word "understand" in a more organic/alive sort of way. While others use it in more of a general logic sense.
eg, if you create a software app that has all the formulas and input/output boxes for doing your taxes does that app now understand tax because the programmer coded understanding into it?  Some people will argue yes, others no.
A lot of disagreements on the internet come down to people having slightly different internal definitions of words

As Dijkstra famously put it  https://amturing.acm.org/award_winners/dijkstra_1053701.cfm "The question of whether computers can think is like the question of whether submarines can swim;"
and in https://www.cs.utexas.edu/~EWD/transcriptions/EWD08xx/EWD898.html
"The Fathers of the field had been pretty confusing: John von Neumann speculated about computers and the human brain in analogies sufficiently wild to be worthy of a medieval thinker and Alan M. Turing thought about criteria to settle the question of whether Machines Can Think, a question of which we now know that it is about as relevant as the question of whether Submarines Can Swim."

[eutectique]

After I corrected it several times, the text became correct, but ...

I noticed a surge in "prompt engineer" jobs over the last year.

[HuronKing]

ChatGPT can be useful when instead of asking it an open-ended question I ask it about a specific error or problem I am having.

I'm starting to use LaTeX and I was having issues with some compiling errors. Googling the compile errors turned up absolutely nothing immediately helpful (I don't have time or patience to scroll through pages and pages of back and forth on Stack Exchange to get the answer). I asked ChatGPt, it gave me 3 clearly worded suggestions for how to fix my problem.

The second suggestion worked.

If you ask ChatGPT a limited scope question with a verifiable way to test its answer, it's a very powerful tool.