OpenAI、Claude、Gemini 隐藏思维链被攻破：防蒸馏防线全面崩塌
三巨头 API 密码学旁路漏洞被证实：轻量模型可作"解密器"恢复旗舰模型隐藏思维链，闭源模型最引以为傲的护城河失效。拆解漏洞三层结构与防蒸馏影响。
英伟达开源的真正目的：AI竞争已从"最大模型"转向"最优路由"
英伟达连发 Nemotron 3.5 Lightning 与 NeMo Switchyard：AI 竞争已从"最大模型"转向"最优路由"，企业部署聚焦高频 Agent 任务与模型调度。
Nvidia's Open-Source Gambit: Why Model Routing, Not Model Size, Is the New Battleground
英伟达开源棋局：模型路由而非模型规模才是新战场。Nemotron 3.5 Lightning（30B MoE）+ NeMo Switchyard 路由库双发布，表面是加入开源竞赛，实则是战略布局。
NeMo Switchyard 上手：给智能体加个模型调度员，成本降 74%
上手英伟达 NeMo Switchyard 模型路由库 + Nemotron 3.5 Lightning：给 Agent 执行层加调度员，按准确率/延迟/成本动态路由，成本最多降 74%。
NeMo Switchyard: Route Agent Workloads, Cut Costs 74%
Long-running AI agents burn most tokens on routine execution. NVIDIA's NeMo Switchyard routes workloads to the right model class — cutting costs up to 74% without sacrificing quality.
Route LLM Traffic with NVIDIA NeMo Switchyard
One model for everything no longer makes sense. Coding agents spend most tokens on mechanical steps. This guide shows how to route LLM traffic with NVIDIA NeMo Switchyard.
NeMo Switchyard 上手：给 LLM 调用加一个智能路由
开发者向：NeMo Switchyard 是 Apache 2.0 的 Rust 路由代理，夹在 Agent 与模型后端之间做协议翻译、类型化路由与可观测性，两分钟跑上智能路由。
Meta Returns to Open Source: Muse Glimmer 30B Runs Locally
Meta placed its biggest bet yet on the open-vs-closed AI debate: Muse Glimmer, a 30B-parameter Apache 2.0 model built for local agentic work, released August 10 with a Zuckerberg manifesto.
Meta 重回开源：30B 本地智能体模型 Muse Glimmer 发布，扎克伯格万字檄文力挺蒸馏
Meta 发布 30B 本地智能体模型 Muse Glimmer（Apache 2.0，单张消费级 GPU 可跑），扎克伯格万字檄文为蒸馏正名，并宣布设独立董事会审核每次模型发布。
DeepSeek 下场造 Harness：只卖模型的时代结束了
DeepSeek 注册"Harness 团队"公众号，正式下场做模型之外的工程系统、对标 Claude Code。Model + Harness = Agent——只卖模型的时代结束了。
DeepSeek Harness: The End of the Model-Only Era
DeepSeek quietly registered its "DeepSeek Harness" team — one of the clearest signals that AI's center of gravity has moved. The lab that built its reputation on models is now building the layer around them.
品牌案例拆解：DTC 护肤品牌如何用 AI 创意测试把广告成本砍掉 40%
DTC 护肤品牌用 AI 创意测试把广告成本砍掉 40%：AI 文案变体 + 程序化测试 + 72 小时淘汰标准，创意迭代提速 20 倍，附完整方法与踩坑复盘。