智能体破解多义性；Meta 发布 Muse Spark Open
原标题：[AINews] Agents disprove Polysemanticity; greatly improve Interpretability & Alignment
AI 摘要
Goodfire 的 Nick Cammarata 发表文章提出「神经网络已经是干净的」理论，认为多义性和叠加现象可通过进一步的智能体分析化解，从而提升可解释性与对齐，并可能降低 P(doom)。Meta 发布 Muse Spark Open，这是一个 816B 总参数/42B 激活的 MoE 模型，支持 1M token 上下文和多模态输入，在 Artificial Analysis 智能指数上得 39 分，为美国开源权重模型最高，但落后于 GLM-5.3 和 Kimi K3。
正文节选
A version of Muse Spark Open was released today, fulfilling their original promise but with a model weaker than their Spark 1.3. It’s also worth watching the Tesla Roadster demo - not strictly AI but nothing cooler than real life flying cars. Our title story today is a rare one on interpretability and alignment. We last caught up with Goodfire in Feb: And now Nick Cammarata at Goodfire has an important essay out stating his working theory that “Neural Networks Are Already Clean” — showing that m