Doing numbers on Twitter/X this week: this paper tested 70+ LLMs on open-ended prompts and found they all produce strikingly similar outputs. Worse, the systems used to improve models actively penalize diversity, reinforcing the convergence.