• Public

  • Public
  • Network
  • Groups
  • Popular
  • People

Image of a very crude lavender-colored quadruped, composed of elementary graphical shapes with a simple horn-like appendage emerging from the head end. Text: The researchers thought GPT-4 might have somehow memorized code for drawing a unicorn from its training data, so they gave it a follow-up challenge: They altered the unicorn code to remove the horn and move some of the other body parts. Then they asked GPT-4 to put the horn back on. GPT-4 responded by putting the horn in the right spot. GPT-4 was able to do this even though the training data for the version tested by the authors was entirely text-based. That is, there were no images in its training set. But GPT-4 apparently learned to reason about the shape of a unicorn’s body after training on a huge amount of written text.