Taylan (Male Feminist Arc) on Nostr: Don't unspoiler the second image before reading, if you want to take a guess about ...
Don't unspoiler the second image before reading, if you want to take a guess about ChatGPT's reasoning skills.
So, this is very interesting...
To prove that LLMs can't actually reason, Sabine points out that ChatGPT makes a rookie mistake and says the smallest integer whose square is between 5 and 17 is 3, when the correct answer is -4 (minus four).
Now, I thought, "but that's a very human mistake! It simply forgot to consider negatives!"
So, I reproduced the problem in a ChatGPT session of my own. It also said 3 to me, like in her video, so points for consistency. See first attachment.
But then I pointed out to it that it forgot to consider negatives. Can you guess whether it then got to the right answer?
Second attachment reveals the answer.
This is GPT-4 by the way. It says it's the October 2023 version.
https://www.youtube.com/watch?v=TpfXFEP0aFs
So, this is very interesting...
To prove that LLMs can't actually reason, Sabine points out that ChatGPT makes a rookie mistake and says the smallest integer whose square is between 5 and 17 is 3, when the correct answer is -4 (minus four).
Now, I thought, "but that's a very human mistake! It simply forgot to consider negatives!"
So, I reproduced the problem in a ChatGPT session of my own. It also said 3 to me, like in her video, so points for consistency. See first attachment.
But then I pointed out to it that it forgot to consider negatives. Can you guess whether it then got to the right answer?
Second attachment reveals the answer.
This is GPT-4 by the way. It says it's the October 2023 version.
https://www.youtube.com/watch?v=TpfXFEP0aFs