AI Fails on Theory of Mind Tasks
Data Skeptic
00:00
How to Get the System to Trip Up
"I introduced four variations to the Smarties task," he says. "And for each one, there were two completions where I was just following in the footsteps of the previous paper" He adds that GPT 3.5 didn't trip up all 16 out of its variants.
Play episode from 39:44
Transcript
