## 🔍 Core Insights This week in AI, **technology deployment is accelerating** while **industry anxiety persists**. On one hand, the **Fudan team** reviewed the development process of the **744B-parameter MoE model** Atria Dawn Preview, showing that the median number of execution operations handled by AI agents jumped from 11 to 28.5 within four weeks [0]; **Alibaba DAMO Academy** open-sourced the abdominal CT diagnostic multimodal model **RADAR** [17], while **Interfaze** launched the efficient quality inspection model **lev** based on **Qwen-3.5 4B** [18]. On the other hand, some early **Anthropic** employees were reported to be considering buying land in remote areas due to fears of "AI going out of control" [7][12], reflecting deep-seated anxiety within the industry about **AGI safety**. ## 🚀 Key Updates - **Fudan team reviews Atria Dawn Preview development** [0]: 56 participants and 700+ task logs show that the median agent operations per person rose from 11 to 28.5 within four weeks, with humans still leading decision-making. - **Alibaba DAMO Academy open-sources RADAR medical AI** [17]: A multimodal model assisting abdominal CT diagnosis, with lesion analysis capabilities approaching expert level. - **Interfaze open-sources lev quality inspection model** [18]: Based on Qwen-3.5 4B, it directly outputs results through internal probabilities, enabling efficient multi-task quality inspection. - **Anthropic senior employees reported to be preparing "AI out of control" shelters** [7][12]: WSJ reports that several early employees are considering buying land in remote areas, and internal Slack once simulated doomsday scenarios. - **GPT-6 completes long-chain WeChat tasks without an interface** [3][11]: Fu Sheng tested GPT-6 continuously executing 120+ WeChat greeting tasks and believes AGI may have already arrived. - **Opus 5.5 and Sol spark team selection disagreements** [8]: Both models have their supporters, and the article also discusses the trend of evals increasingly becoming a hiring requirement. - **Privatemode demonstrates GLM-5.3-Flash typed decision method** [21]: Without fine-tuning, it reads numbered options and log probabilities to output decisions with probabilities in a single forward pass. - **Nubia NaviX Ultra users forced offline while playing Honor of Kings** [4]: Doubao mobile assistant stated it did not perform any operations on Tencent's game system and advised users to temporarily stop trying to log in. ## 🔗 Sources [0] Fudan team reviews Atria Dawn Preview development: AI agents take on more execution, humans still make the vast majority of decisions — https://aihot.news/items/cmujz82l112d4ro9hsb2680j6 [3] AI long-chain task succeeds in WeChat environment without an interface — https://www.bestblogs.dev/status/2104209485489430644?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item [4] Nubia NaviX Ultra users forced offline while playing Honor of Kings, Doubao mobile assistant says it did not perform any operations on Tencent's game system — https://aihot.news/items/cmujwzb0a1020ro9h28gpqmyg [7] Some senior Anthropic employees consider buying land in remote areas to prevent "AI out of control" — https://aihot.news/items/cmujwzb0b1022ro9h55cz1o8y [8] Opus 5.5 and Sol split teams in two — https://aihot.news/items/cmuj