| ▲ | gwern 4 hours ago | |
> The goal of the game is to give hints to a secret word without giving the same hint as another player; it was shocking how often all the agents playing would give the same hint, even at temperature 1. We resolved this by giving them all “personalities” which were really just topics; for example we’d be telling one agent to think about things in the context of sports, so their clue for “shell” might be “defense” or something, whereas the agent told to be a hippie might say “cancer” for the zodiac’s crab. Mode-collapse is characteristic of pretty much all chatbots post-davinci-002 and a major challenge to anything creative or game-like. | ||