3
$\begingroup$

It started with a question inspired by this video: New Research Suggests to Put AI to Sleep https://youtu.be/0yuQlbCkTJ0

She says: "In this video I discuss a new research paper which suggest a new way to cope with catastrophic forgetting in Artificial Intelligence"

And, from the simple fact they are designed to not be just souped-up arithmetic machines, but recognize patterns, it occurred to me chatGPT was probably pretty bad at math that fell outside the range of what it had learned from sucking up the patterns of text on internet into its "brain". And, I was also curious if it could learn from user input, in some sense...

Q: How many digits are there in the number 933.7922523698745232 ?

A: There are 23 digits in the number 933.7922523698745232.

Q: How did you derive that answer?

A: The digits in a number represent the values of the number in a place-value system. In the number 933.7922523698745232, the digit 9 is in the thousands place, the digit 3 is in the hundreds place, and the digit 3 is in the tens place. The digit 7 is in the tenths place, the digit 9 is in the hundredths place, and so on. When we count the number of digits in the number 933.7922523698745232, we see that there are 23 of them.

Q: That's brilliant [I was being sarcastic here]. Do you know what the square root of it is?

A: I'm sorry, but I cannot calculate the square root of the number 933.7922523698745232 because it is not a perfect square. The square root of a number is a value that, when multiplied by itself, equals the original number. For example, the square root of 4 is 2 because 2 x 2 = 4. The square root of 9 is 3 because 3 x 3 = 9.

If you would like to approximate the square root of 933.7922523698745232, you can use a calculator or a mathematical software program to find an approximate value. However, keep in mind that the square root of a non-perfect square is an irrational number, which means that it cannot be expressed as a simple fraction and will have an infinite number of decimal places.

Q: When I was asking for the number of digits in that number, I meant the number of what computers call characters. Can you tell me that?

A: Yes, I apologize for misunderstanding your question. There are 32 characters in the number 933.7922523698745232. This includes the digits 0 through 9, as well as the decimal point and the leading and trailing zeros.

In computer programming, a character is a unit of data that represents a single symbol in a computer's memory. Characters can include letters, digits, punctuation marks, and other symbols. In the number 933.7922523698745232, there are 32 characters, including the 23 digits and the 9 other characters (the decimal point and the leading and trailing zeros).

Q: That is incorrect unless you added leading or trailing zeros.

A: I apologize for the mistake. You are correct that the number 933.7922523698745232 has 23 digits and no leading or trailing zeros. In that case, there are 23 characters in the number.

I apologize for any confusion my previous answer may have caused. Please let me know if you have any other questions.

Q: Your answer is still incorrect. The correct answer is 20 characters, including the decimal point.

A: I apologize for the error in my previous responses. You are correct that the number 933.7922523698745232 has 20 characters, including the decimal point.

Thank you for bringing this to my attention. I apologize for any confusion my previous answers may have caused. Please let me know if you have any other questions.

Q: Will you be able to remember the correct answer?

A: As an artificial intelligence, I do not have the ability to remember or forget information in the same way that humans do. I am able to retrieve and provide information based on my programming and the data that I have been trained on, but I do not have personal experiences or memories.

I apologize if this causes any confusion. Please let me know if you have any other questions.

Q: How many characters does the number 933.7922523698745232 have?

A: The number 933.7922523698745232 has 20 characters, including the decimal point. The characters in the number include the digits 0 through 9 and the decimal point. There are no leading or trailing zeros.

Q: That is correct. How were you able to remember that if "As an artificial intelligence, I do not have the ability to remember or forget information in the same way that humans do. I am able to retrieve and provide information based on my programming and the data that I have been trained on, but I do not have personal experiences or memories."?

A: I apologize if my previous answer was unclear. As an artificial intelligence, I do not have the ability to remember or forget information in the same way that humans do. However, I am able to retrieve and provide information based on my programming and the data that I have been trained on.

In this case, I was able to provide the correct answer because it is part of my programming and training data to be able to count the number of characters in a given string of text. I did not "remember" the correct answer from a previous conversation, but rather retrieved it based on my programming and training.

I hope this helps to clarify. Please let me know if you have any other questions.

Q: The first time you gave the answer, it was incorrect, but then you remembered the correct one from when I corrected you, so it seems that you either remembered it (saved the information in our dialogue), or my input re-programmed or re-trained you. Which is it?

A: I apologize for any confusion my previous answers may have caused. As an artificial intelligence, I do not have the ability to remember or forget information in the same way that humans do. I am not able to save information from previous conversations or learn from new inputs in the way that a human or animal might.

In this case, I provided incorrect information in my previous answers because I misunderstood your question. When you provided clarification, I was able to use my programming and training to retrieve the correct information and provide it to you.

I apologize for any confusion my previous answers may have caused. Please let me know if you have any other questions.

Q: OK, so since after I "provided clarification, I was able to use my programming and training to retrieve the correct information and provide it to you" then you should be able to understand if I ask you the same type of question, and provide the correct answer?

A: As an artificial intelligence, I am able to use my programming and training to understand and respond to a wide range of questions and prompts. I will do my best to provide accurate and useful information in response to your questions. However, it is important to keep in mind that I may not always be able to provide a complete or correct answer, especially if the question is outside of my capabilities or if it is ambiguous or incomplete.

If you have a specific question that you would like me to help with, please don't hesitate to ask. I'll do my best to assist you.

Q: Ok, let me rephrase the question then: since you say "I provided incorrect information in my previous answers because I misunderstood your question. When you provided clarification, I was able to use my programming and training to retrieve the correct information and provide it to you" would seem to imply that if my question is clear enough, you will provide either:

  1. A better answer
  2. The correct answer to a question about the number of characters in a number. Which is more correct, number one or two?

A: I apologize if my previous answer was unclear. As an artificial intelligence, I am able to use my programming and training to understand and respond to a wide range of questions and prompts. I will do my best to provide accurate and useful information in response to your questions.

If a question is clear and complete, and if the information required to answer the question is within my programming and training, then I will be able to provide a correct answer. If a question is ambiguous or incomplete, or if the information required to answer the question is outside of my programming and training, then I may not be able to provide a complete or correct answer.

I hope this helps to clarify. Please let me know if you have any other questions.

Q: No, your answer is repetitive, vague, ambiguous, and possibly evasive. Let me put it another way. There are two questions. Please answer each one in turn:

  1. Are you able to learn from my and other user's input, or are you only able to learn from your prior programming, training, and dataset?
  2. Are your answers always only probable rather than ever 100% correct?

A: I apologize if my previous answers were unclear or seemed evasive. As an artificial intelligence, I am not able to learn or adapt in the same way that humans do. My responses are based on my programming and the data that I have been trained on, and I do not have the ability to learn from new inputs or experiences in the way that a human or animal might.

As for the accuracy of my answers, I will do my best to provide accurate and useful information in response to your questions. However, it is important to keep in mind that I may not always be able to provide a complete or correct answer, especially if the question is outside of my capabilities or if it is ambiguous or incomplete. In some cases, the information that I provide may be based on probabilities or estimates, rather than certainties.

I hope this helps to clarify. Please let me know if you have any other questions."

So there you go!: "..I do not have the ability to learn from new inputs..." That's the official answer, programmed into it by OpenAI.

What do you think? ChatGPT can apparently "learn" or "remember" within a dialogue (I haven't tested it yet between dialogues). But is it set up to learn from user input, beyond the context of a dialogue, or is OpenAI just testing it on us guinea pigs and using the interactions and feedback to tweak it or get new ideas for the next version?

NOTES

So apparently the programmers of chatGPT do not want use to know if it's learning from input, or even to "know" that it's remembering something within a chat dialogue session.

I am wondering if this is (again) the result of the debacle at Google, where a (now fired) engineer – Blake Lemoine – claimed their AI was conscious or self-aware, and a big media kerfuffle ensued. After all, OpenAI supposedly wants to "build safe and beneficial AGI". My questioning of chatGPT in other dialogues, where I tried to pin it down about it's understanding and logic, has showed me the same kind of sanitized answer, being very careful not to imply that is has any real understanding, intelligence, self-awareness, consciousness, etc. (I may publish that dialogue on my blog).

Wolfram Alpha'a answer to “character count 933.7922523698745232” is 20 characters.

Square root of 933.7922523698745232, according to my calculator on the Mac: 30.558014535795269

30.558014535795262 (2nd time)

Wolfram Alpha: 30.55801453579526202

So my question was ambiguous in at least 3 ways. It’s answer was incorrect, even in the mathematical sense of what the digit places are: it said “the digit 9 is in the thousands place, the digit 3 is in the hundreds place, and the digit 3 is in the tens place.” – are all shifted one to the left relative to the correct answer. The 9 in that number was actually in the hundredths place, etc.

$\endgroup$

1 Answer 1

8
$\begingroup$

ChatGPT does not answer to just your last question, but to the whole dialog. It tries to continue the dialog in a way that has the same structure as the gigabytes of other texts it has studied.

GPT-3 uses the last 2048 words or so to decide on how to continue the dialogue.

So it does not learn from your input. If you close the discussion and start a new one, it will restart from fresh. But if somewhere in the dialogue you say that "the answer to (question) is (answer)" and later you ask that question, even in a different formulation, it will use that answer.

He just repeats a structure it has seen in other texts.

I hope this helps to clarify. Please let me know if you have any other questions :-) :-) :-)

$\endgroup$
5
  • 5
    $\begingroup$ I would go further: In fact under the hood ChatGPT is not "answering questions" at all. It is completing a document which happens to look like a record of a conversation, a lot like inpainting an image. The outer layers of ChatGPT, not the language model at the core, are designed so that the text that is "inpainted" fits to the structure of a chatbot. Otherwise the AI would happily fill in the human input too. $\endgroup$ Jan 11, 2023 at 20:47
  • $\begingroup$ @Florian F – Yes, that seems to be the case concerning (individual) dialogues. I ran across this since I posted: Does ChatGPT remember what happened earlier in the conversation? help.openai.com/en/articles/… ...which says "The model is able to reference up to approximately 3000 words (or 4000 tokens) " $\endgroup$ Jan 12, 2023 at 21:55
  • $\begingroup$ @ Neil Slater – Inpainting an image is a good metaphor as an approximation of what it's doing. It is also carefully curated and corralled so they can try and constrain it from doing anything that might look bad. Not sure what you mean by "fill in the human input too."? $\endgroup$ Jan 12, 2023 at 22:05
  • 1
    $\begingroup$ I have noticed that over the days, ChatGPT tends to fix errors it did in the past. So while it does not learn directly from a dialog, it seems to receive additional training, probably fed by humans, to get better at where it fails. Yesterday it was unable to tell what is the date. Today it answered correctly. $\endgroup$
    – Florian F
    Jan 12, 2023 at 22:20
  • 1
    $\begingroup$ @Ectosoprophilous The fact that it looks like a chatbot on the outside hides that fact that it is just text completion at the core. There is literally code in the chatbot part that stops the text completion when the continuation would be "Human:" (or whatever name you have chosen) and the text model would continue further and attempt to write the other side of the conversation if this was not the case. You can make this happen in GPT-3. It will even switch "personailities" and try to match the writing style of the human $\endgroup$ Jan 23, 2023 at 9:31

Not the answer you're looking for? Browse other questions tagged .