Algorithmic tacit collusion — LessWrong