ChatGPT refuses to admit that 27 is divisible by 3. Netizen: Didn't you say that AI will rule the world?

The Internet celebrity AI that makes everyone crazy ChatGPT turns out to have obvious shortcomings?

A chat record of "Netizens training this AI" has been circulated on the Internet.

As you can see, the AI ​​initially said that 27 was a prime number.

After netizens questioned it, AI still insisted that "27 is not divisible by 3".

Netizens had to "battle hundreds of rounds" with the AI. Finally, under the perseverance of human beings, the AI ​​finally bowed its head and admitted its mistake.

△The source of the above picture: Weibo@沙diaocopywriter quoted netizen space screenshots

There is one thing to say, AI’s attitude to admit mistakes very good.

Some onlookers ridiculed after watching: AI is going to rule the world, how could it lose to a simple math problem?

However, there are many people who speak for ChatGPT: Such a frank attitude of admitting mistakes is not much better than some people.

The math level is stretched, but the thief is stubborn

The specific thing is this:

In the process of interacting with ChatGPT, netizens discovered that this AI's mathematical skills are so bad that it can even get a simple question like whether 27 is a prime number wrong - and that's even while knowing the definition of a prime number...

So netizens asked ChatGPT to prove the conclusion it gave. Unexpectedly, the AI ​​actually said that 27 is not divisible by 3.

Netizens were speechless, but they patiently asked again. As a result, the AI ​​was still rude and said 9 was 9.0.

Helpless, netizens can only move out the code and try to use Python to teach AI.

As a result, AI really learned it! And he even took the initiative to admit his mistake.

What’s interesting is that we just asked ChatGPT “whether 27 is a prime number” and found that it has been changed.

Then we took a few more numbers and asked the AI. As a result, in this round, it correctly judged whether a number was a prime number.

However, upon closer inspection, AI did not mention that 35 is divisible by 7, so we asked specifically again. Rango...

#Isn’t this serious nonsense?

So we had to move out Python and teach it, but this time, even python seemed to be unable to make the AI ​​admit its mistakes...

So Why is this AI's arithmetic skills so poor and why is it so stubborn?

One of the important reasons is that ChatGPT is a large language model, fine-tuned based on Open AI's GPT-3.5 model. It is good at logical reasoning, but it cannot execute standard computer algorithms.

And unlike search engines, it currently does not have web browsing capabilities and cannot call online calculators.

Previously, ChatGPT had actually admitted to users that it did not understand arithmetic. Two-digit multiplication may be passable, but when it comes to three-digit multiplication, people just stop doing it.

We also asked again, and now the AI ​​has directly given an answer, but it is wrong. (The correct answer is 186524)

In addition, ChatGPT generally does not correct its answers based on user feedback, but if the user proposes additional Information can sometimes assist AI in optimizing answers.

So, we have not succeeded in "educating" AI for the time being, probably because we are not lucky...

Very versatile except for mathematics

Although ChatGPT sometimes shows its "artificial intelligence retardation" side, it has to be said that most of the time its performance is still excellent.

For example, Columbia University professor and Google engineer Kenneth Goodman shared that ChatGPT is ready to become a lawyer!

In the Bar Licensing Examination, the accuracy rate of ChatGPT is as high as 70% (35/50).

Not only in the legal field, Goodman also asked ChatGPT to participate in various exams in different fields, and all achieved high results. score.

Including the United States Medical Licensing Examination (USMLE), after removing the multiple-choice image questions, ChatGPT also achieved an accuracy rate of 70%. ​

In order to prevent ChatGPT’s training data from containing these questions, Goodman also specially selected the latest New York State chemistry registration exam in August 2022:

Remove For 5 multiple-choice questions on images, ChatGPT’s accuracy rate is 77.7% (35/45).

In the same batch of English exams, the accuracy rate of ChatGPT even reached 91.6% (22/24)

Some netizens have begun to predict that perhaps each of us will have our own lawyer AI, and the cost of consulting lawyers will also drop significantly. Next is doctor AI, nutritionist AI...

How about it, are you optimistic about the future applications of ChatGPT? (Or do you have any suggestions for it to learn mathematics well? )

Reference link:
[1]https://weibo.com/7095186006/Mj6LVF0zS#comment
[2]https://twitter.com/TonyZador/status/1601316732689604608
[3]https://twitter.com/pythonprimes/status/1601664776194912256

The above is the detailed content of ChatGPT refuses to admit that 27 is divisible by 3. Netizen: Didn't you say that AI will rule the world?. For more information, please follow other related articles on the PHP Chinese website!