According to @godofprompt on X, Google researchers report that frontier reasoning models like DeepSeek R1 and QwQ32B exhibit spontaneous internal multi-agent debate within their chain of thought, emerging from reinforcement learning for accuracy rather than explicit training, and that amplifying this multi-perspective dialogue further improves performance on hard tasks. As reported by @godofprompt, the study argues that longer chain-of-thought alone does not yield better results; instead, distinct internal perspectives that question, verify, and contradict one another causally account for gains, a phenomenon the authors call a society of thought. According to @godofprompt, the business implication is that future AI systems should adopt organizational design patterns—roles, norms, and protocols—similar to courtrooms and markets, moving beyond single-threaded transcripts to structured disagreement for higher reliability and scalability.