GPT-4 performs better at Theory of Mind tests than actual humans

Re:GPT-4 performs better at Theory of Mind tests than actual... by CamperBob2 on Sunday April 30, 2023 @06:43PM (#35767239)

Mimicry is enough to account for the output, but it doesn't explain how the model is so good at understanding even vague or poorly-considered prompts.

"Autocompletion" isn't an adequate explanation when neither the prompt nor the output has ever existed before. Something else is going on... something that obviously has a lot in common with how our own minds work.

[ Parent | on HN ]

Re:GPT-4 performs better at Theory of Mind tests than actual... by fl7305 on Sunday April 30, 2023 @07:39PM (#35767566)

I agree. Another way to come to this conclusion is that the GPT-4 model is on the order of 1000B parameters, or on the order of a terabyte of data.

That's not a lot of different sentences when you consider the combinatorical explosion.

[ Parent | on HN ]