I know 4chan is being cynical on this topic, but GPT is actually, unironically amazing. It can solve simple arithmetic problems, it can translate language with minimal priming, it can fucking REASON. All of that just from being trained to predict what words should come after other words. It somehow gains fundamental insight on mathematics, logic and translation from that. It's mind boggling to me. Here's a few examples from GPT-2 (Griffon Model)
>If I put my shoes on my hands and do a handstand what are my shoes touching?
"The ground."
>Please explain your reasoning
"Because your hands are on the ground, your shoes would be touching the floor."
>What color paint would I have if I mixed paint that is the color of the sky and paint that is the color of a dandelion?
"You would have a strange color, probably greenish." (Third completion)
>Please explain your reasoning
"The first paint would be blue. The second paint would be yellow. If you mixed the two paints, you might get a strange color, probably greenish." (First completion)
>There is a ball inside a box. The box has holes in it. The holes are too small to fit your arm through, but large enough for the ball to fit through. How do you get the ball out of the box?
"I would use my fingers to push the box around so the ball would fall out." (Third completion)
>Please explain your reasoning
"Because I am looking at the box, the ball is inside of it. The holes are big enough for the ball to roll out of the box. Because the holes are big enough, I can move the box around so the ball will roll out of it." (Eighth completion)
>A giraffe and a donkey are playing catch when their ball lands in a tree. Please describe how the donkey might get the ball.
"The giraffe would reach up and grab the ball out of the tree."
>Please explain your reasoning
"Because the giraffe is a tall animal, he would be able to reach the ball. The donkey wouldn't have any such ability, and would need the help of another animal." (First completion)
