AI 可解释性接近突破？一篇新文章让研究者意见分裂
原标题：Is AI interpretability close to a breakthrough? A new essay has researchers divided
AI 摘要
Goodfire 研究员 Nick Cammarata 发表长文《Neural Networks Are Already Clean》，推翻自己早年参与提出的叠加假说（superposition），认为神经网络内部比学界此前认为的更有条理，并给出 Qwen3-32B 中约 90% 神经元极点有单一主要功能的初步数据。OpenAI 的 Roon 称这可能是科学史上的重大时刻，而 Google DeepMind 的 Neel Nanda 则警告不要过度炒作，认为这只是有用的工具而非突破。若结论成立，将利好 AI 安全与对齐研究。
正文节选
Is AI interpretability close to a breakthrough? A new essay has researchers divided Neural networks aren't nearly as tangled as researchers assumed, according to Nick Cammarata, one of the people behind the superposition hypothesis. OpenAI's Roon calls it a historic moment. Google DeepMind researcher Neel Nanda warns against overhyping it. In a long essay titled "Neural Networks Are Already Clean," Nick Cammarata argues that the inner workings of large AI models are much easier to understand tha