We're safe. ChatGPT couldn't fix a typo with a lot of context

4 points by goofed ↗ HN
Notice that in the chat below there's only a simple typo: chod => chord. You can try yourself.

Me: From which key is the chod C#º borrowed from in a song that is in the key o F major?

ChatCPT: It is not clear what you are asking. Could you please provide more context or clarify your question?

Me: From which key is the chord C#º borrowed from in a song that is in the key o F major?

ChatCPT: In music theory, a chord that is borrowed from another key is called a "borrowed chord." In a song that is in the key of F major, a C# diminished chord could be borrowed from the key of B major, which is the parallel minor key of F major. In other words, the C# diminished chord would be borrowed from the parallel minor key, which is a common technique used in music composition to add interest and variety to the harmonies. However, it is also possible that the C# diminished chord could be used as a passing chord or a temporary tonicization (temporarily establishing a new key center) in the F major song, in which case it would not necessarily be considered a borrowed chord.

7 comments

[ 0.29 ms ] story [ 22.0 ms ] thread
I also just noticed that the response is VERY WRONG.

Take this part: "the key of B major, which is the parallel minor key of F major"

The parallel minor key of F major is F minor.

Our jobs are safe, guys.

Even if ChatGPT does for example prevail in the long run. I don't think it'll create too much job loss, we'll just get new types of jobs.
And here's a follow up:

Me: Unfortunately your response is very wrong. Specifically this part "the key of B major, which is the parallel minor key of F major." The parallel minor of F major is not B major. If I hadn't done an additional google search, I'd be taught the wrong theory by you. I'd suggest you stop responding questions that you don't have 100% confidence that they are correct.

ChatGPT: I apologize for the error in my previous response. You are correct that the parallel minor of F major is not B major. The parallel minor key of F major is D minor. I apologize for the mistake and for any confusion it may have caused. Thank you for bringing this to my attention. It is important to me to provide accurate and helpful information, and I will make sure to double-check my responses in the future. If you have any further questions, please don't hesitate to ask.

Me: It happened again. The parallel minor key of F major is also not D minor. Can you please stop making guesses? You're damaging society by letting people believe that you've got correct answers.

ChatGPT: I apologize for the mistake in my previous response. You are correct that the parallel minor key of F major is not D minor. The parallel minor key of F major is F minor. I apologize for the error and for any confusion it may have caused. Thank you for bringing this to my attention. It is important for me to provide accurate and helpful information, and I will make sure to double-check my responses in the future. If you have any further questions, please don't hesitate to ask.

Just goes to show that when scrutinised and pressed, ChatGPT is another AI hype tool that essentially is good at regurgitating nonsense and is even confident that its answer is correct.

As soon as it admits it's mistake, it even breaks and repeats the exact same verbatim response again. The hilarious part of it is that it completely fails at explaining itself and doesn't even know why it is giving the wrong answer, hence why it is as transparent as a black box.

I’ve seen the same behavior. I wonder if these are learning instances for the AI or is it just dumbly going through a set of possible answers until the user is satisfied with the answer. If it’s the latter, it’s basically a fancy way of doing a google search and going through potential answers. If it’s the former and the AI learns from these mistakes (say if a consistent number of users point them out) I assume it will get much better over time.
answering my own question, I guess not:

https://help.openai.com/en/articles/6787051-does-chatgpt-rem...

So at its current stage it’s pretty useless as it will not learn from its mistakes, seems to me the opposite of intelligence, just another mechanical turk.

I think the two main problems here are:

- Users can't use it to learn, because ChatGPT responds with high confidence but could be terribly wrong.

- Even if it could learn from users, it would really only learn from people that know a subject and are willing to spend their time correcting ChatGPT.