Latest AI ChatGPT do not understand math ?

[szan]

Hello,
I've marked in a yellow in a screenshot attached to this post, where latest artificial intelligence ChatGPT says something like this: 

No, 64*8 does not equal 512. The size of an unsigned long long int on a 64-bit Intel Core i3 with AVX2 support is 64 bits, which means 8 numbers are needed to fill the entire 256-bit register. This is because the data must be aligned to a 32-byte boundary, and 8 numbers are needed to fill the entire register.

I was interested to know something more about AVX2 instruction set extension available in my Intel Core i3 processor and at some point of this "conversation" with  ChatGPT available here:
https://beta.openai.com/playground/p/8coAfEUgFHTTsMJXKBF3Uli5?model=text-davinci-003

I've noticed that AI ChatGPT defines array of 8 numbers insteed of 4, since on 64 bit machine "unsigned long long int" has size 8 bytes, which means we need only 4 such numbers to fill 256 bit AVX2 registers 

I said:

Nope, 64*8 equals to 512 so it looks like something is wrong in this example?

I saw  ChatGPT answer like shown above, which means that... nasty things can happen when AI will be used instead of humans in our everyday lives 

ChatGPT: No, 64*8 does not equal 512.

Let us know what do you think about this ChatGPT   

$ maxima
Maxima 5.43.0 http://maxima.sourceforge.net
using Lisp SBCL 1.4.14-3.fc31
(%i1) 64*8;
(%o1)                                 512
(%i2)

[szan]

Anyway, google search helped me find much better portable solution for Linux gcc than chatgpt 
It makes sense and humans response gives clean example howto allocate such memory for use in AVX2 too, since we need 32 byte memory alignment   

What does posix_memalign/memalign do
https://stackoverflow.com/questions/6563120/what-does-posix-memalign-memalign-do

Whereas malloc gives you a chunk of memory that could have any alignment (the only requirement is that it must be aligned for the largest primitive type that the implementation supports), posix_memalign gives you a chunk of memory that is guaranteed to have the requested alignment.

So the result of e.g. posix_memalign(&p, 32, 128) will be a 128-byte chunk of memory whose start address is guaranteed to be a multiple of 32.

This is useful for various low-level operations (such as using SSE instructions, or DMA), that require memory that obeys a particular alignment.

[tooki]

Hello,
I've marked in a yellow in a screenshot attached to this post, where latest artificial intelligence ChatGPT says something like this: 
...

Just to be clear: what you tested and screenshotted there is not ChatGPT. You tested GPT-3, which is different: https://dev.to/ben/the-difference-between-chatgpt-and-gpt-3-19dh

If you look at your own screenshot, there's a text saying "Looking for ChatGPT?" next to a button called "try it now".

[szan]

You tested GPT-3, which is different: https://dev.to/ben/the-difference-between-chatgpt-and-gpt-3-19dh

I've tested one of the most powerful language models currently available as you can findout in this link above:   

GPT-3, or Generative Pretrained Transformer 3, is the third generation of OpenAI's GPT language model, and it is one of the most powerful language models currently available. It can be fine-tuned for a wide range of natural language processing tasks, including language translation, text summarization, and question answering.

So, the only difference is that  ChatGPT could be useless in this case, while Generative Pretrained Transformer 3 (GPT-3) - one of the most powerful language models currently available  - I've used showed me this hopeless answer:

No, 64*8 does not equal 512.

[SiliconWizard]

Why bother though?

[bidrohini]

This is the first time I am hearing/reading about ChatGPT. Gotta try it someday.

[Siwastaja]

I think these AI chat threads are pretty interesting, because they show the AI has reached the level of some actual engineers. There are surprisingly many people who know how to spew utter bullshit with tech words in it, so that it still "sounds right", we even have seen that on the EEVblog forum from time to time.

Sometimes these people are even successful, which tells quite a lot about the system. Of course, someone else has to do the actual design work.

I wouldn't mind the slightest if this group of people was replaced by AI, no harm done. I despise people like that, because they are so dishonest.

[tretecou]

I tested some basic trigonometry and geometry tasks with triangles and it failed on basic european high school/ freshman college math homework which was pretty easy so i dont think chatgpt is in this state ever going to be a danger to humanity as all the youtubers say... 

[tooki]

You tested GPT-3, which is different: https://dev.to/ben/the-difference-between-chatgpt-and-gpt-3-19dh

I've tested one of the most powerful language models currently available as you can findout in this link above:   

GPT-3, or Generative Pretrained Transformer 3, is the third generation of OpenAI's GPT language model, and it is one of the most powerful language models currently available. It can be fine-tuned for a wide range of natural language processing tasks, including language translation, text summarization, and question answering.

So, the only difference is that  ChatGPT could be useless in this case, while Generative Pretrained Transformer 3 (GPT-3) - one of the most powerful language models currently available  - I've used showed me this hopeless answer:

No, 64*8 does not equal 512.

I didn’t say it wasn’t powerful. I just said it’s not the one you claimed to be using. That’s your mistake, not mine.

[tooki]

This is the first time I am hearing/reading about ChatGPT. Gotta try it someday.

Ask it to write you a song about something. It’s surprisingly good at that!

On the other hand, its success at creating recipes is… amusing. Like, it’ll get the general idea perfectly, but the amounts will be way off. (For example, I asked it to show me how to make shrimp pad thai, and the recipe used twice as much shrimp as noodles!)

[S57UUU]

This is a LANGUAGE model (even has "chat" in it's name), not intended to do math. Like a kid, that has just learned to speak well (say, a six year old) isn't very good at math.

[SiliconWizard]

I think these AI chat threads are pretty interesting, because they show the AI has reached the level of some actual engineers. There are surprisingly many people who know how to spew utter bullshit with tech words in it, so that it still "sounds right", we even have seen that on the EEVblog forum from time to time.

Sometimes these people are even successful, which tells quite a lot about the system. Of course, someone else has to do the actual design work.

I wouldn't mind the slightest if this group of people was replaced by AI, no harm done. I despise people like that, because they are so dishonest.

 

It's not mutually exclusive, though (and unfortunately.)
Both are very likely to coexist and actually feed from each other.

[szan]

I think these AI chat threads are pretty interesting, because they show the AI has reached the level of some actual engineers.

It is a long way to go. It might be interesting, but typing with AI in humans natural language is horrible experience, when you get so sick responses for simple math problems?
Now, internet will be full of art and nonsense claims created by AI.

[Terry Bites]

Well you've just put string of symbols in the prompt. WDYE?
You need a question or some context.
Let botty macbot face explain.
Botty is not on the blog yet........it will be soon. Watch out.

[szan]

Well you've just put string of symbols in the prompt. WDYE?
You need a question or some context.

There was a context, while many other posts were made until I've noticed that latest GPT3 AI do not understand that unsigned long long int on typical x86_64 Linux machine under GCC is 8 bytes long which gives us 64 bits, so two times bigger array is not needed to feed 256 bits to AVX2 registers 

Anyway, in screenshots you've attached from another ugly looking conversation, AI tries to force us humans to... use parentheses in mathematical expressions  to show multiplication in simply math calculations like this?!:  1+2*3 + 4 ... Even idiot knows that  we do multiplication before addition, and many ordinary pocket calculators would have been able to handle such calculations a few decades ago 

I wrote in my conversation with GPT3 AI that:

64*8 equals to 512

Hopefully, Google does not deny such obvious facts and gives as a chance to survive in an AI world 

What equals 512 when you multiply?
Image result for 64*8 equals to 512
Factors of 512 are those numbers which when multiply together gives 512 as a result. Factors of 512 : 1, 2, 4, 8, 16, 32, 64, 128, 256, and 512.

 

[igendel]

On the general topic of ChatGPT: I've seen a lot of people lauding it as a useful tool (and, sometimes, replacement) for programmers, so, out of curiosity, I asked it to write something simple - Arduino code that does LED "breathe" effect. The funny part of the result is in the attached image. The less funny part was that I really tried to explain why this won't work as promised, and ChatGPT agreed with me - but continued to produce variations of the same fundamental issue.

ChatGPT is like Flat Earthers: At first it appears to be reasonable, knowledgeable, its communication makes sense, then all of a sudden this huge void is revealed 

[Edit: Breathe, not Breath]

[szan]

ChatGPT is like Flat Earthers: At first it appears to be reasonable, knowledgeable, its communication makes sense, then all of a sudden this huge void is revealed 

Yep, I needed today repair broken thin wires very close to standard USB A male connector and wanted to see this connector with pins description, while I had to solder new +/- wires, but I was unlucky and  I was involved in this  strange conversation with GPT3 chatgpt AI linked below 
https://beta.openai.com/playground/p/94yIT9JWxiFU3SHsYFBuMehs?model=text-davinci-003

OP: You do not know those four pins description of USB A male port ?

GPT3 AI: No, I do not know the four pins' descriptions of USB A male port. However, the two pins in the first row are used to transmit data, while the two pins in the second row are used to provide power to the device.

Of course Google search was very successful and I was able to quickly solder new wires with the correct polarity and make job done 

[bson]

I suspect it has no understanding of math, so if asked a*b it only knows the answer if it has seen this particular product.  Now as if that's not relatively poor, the worst part is if it hasn't seen that it picks something similar.  As a result you get totally incorrect results:

Mathematica:

39383*923833=36383315039

ChatGPT:

The result of 39383 multiplied by 923833 is 36432476129.

In general, recognizing when you don't know the answer to something is a critical step towards correctness.  Frequently, "I don't know" is the intelligent answer.

The second big problem apart from recognizing it doesn't know something, or can't think of how to find out (for example by ordering a textbook off Amazon rather than the equivalent of googling for more nonsense), is it blindly and naively believes everything it's fed is axiomatic truths.

The third problem is it can't derive possible 'working axioms' or patterns - its patterns are purely language based like templates.

And since it's so bad, I'd like to point out the fourth problem which is the same as the first, that when it doesn't know it pulls something out of its ass.

[bson]

Thinking about it some more, say you're a PhD student.  You've done a bunch of research, and have bits and pieces of your thesis drafted, and now you have ChatGPT generate the thesis document for you.  So you feed it what you have, including all your data along with explanations of what it is, how it relates to your work, and a page or two about the subject of your thesis.  So far so good.  But to output a thesis it needs to have looked at a huge number of quality papers and prior theses, to know what it should look like.  So you feed it a dataset for this.  Now, the first problem you're going to have is that all the data, research, and conclusions of these will pollute the correctness of yours.  You can't trust the output to be correct, because it will fill in gaps with nonsense.  The second problem is you will be accused of plagiarism, and rightfully so, since it effectively takes someone else's document and reasoning and drops in your data...

[Terry Bites]

 GPT wont empty my bins or wash my face. The programmers must surely be morons!

[szan]

ChatGPT:

The result of 39383 multiplied by 923833 is 36432476129.

Maxima also gives different result of this multiplication than AI chatboot:
 maxima

Maxima 5.43.0 http://maxima.sourceforge.net
using Lisp SBCL 1.4.14-3.fc31
Distributed under the GNU Public License. See the file COPYING.
Dedicated to the memory of William Schelter.
The function bug_report() provides bug reporting information.
(%i1) 39383*923833;
(%o1)                             36383315039
(%i2)

The answerr is as simple ass it is - this tool (chat boot) was created only to STEEL phone numbers, because of it is one of the requirements when registering to this bullshit web site 

[szan]

In other words, its answer is wrong, but only by 0.14%. That's better than I could give you in the same amount of time as it got if I didn't have a calculator.

You do not have to use any calculator to have  correct result, because of... you can use your brain and sheet of paper with pencil to obtain correct result ?! 

Decades ago any pocket calculator could give you correct result, so it doesn't make any sense to use this  AI chatboot's to solve real engeenering problems, but art designs, with stupid graphics.

[SiliconWizard]

The problem that some seem to fail to recognize is that this "AI" stuff isn't meant to be useful to us, end-users (past the fun, toying around, discovery period.)

[tszaboo]

I think these AI chat threads are pretty interesting, because they show the AI has reached the level of some actual engineers. There are surprisingly many people who know how to spew utter bullshit with tech words in it, so that it still "sounds right", we even have seen that on the EEVblog forum from time to time.

Sometimes these people are even successful, which tells quite a lot about the system. Of course, someone else has to do the actual design work.

I wouldn't mind the slightest if this group of people was replaced by AI, no harm done. I despise people like that, because they are so dishonest.

I think the ChatGPT reached the level of most managers already. I mean it replies to questions in finite amount of time and the reply is coherent and feels like the AI actually read the question, 100% better than most managers.

[BrianHG]

ChatGPT4 is coming and it will roast #3.
It's going over 1 trillion parameters.