💭 gemini seems approximately as bad as sonnet3.5new at social deduction games.

  • game involves a heap of logical reasoning and rules understanding, where one error causes bad reasoning later in the chain
  • many player talking based game means the context window gets filled with bloat, and you have to fact-check all of it
  • strong leaning to think lying implies evil team even when there are good team characters who need to lie to be effective. They never seem to get this or take the time to think from that hypothetical perspective and see that it explains the actions