生成式AI 一、 H20之后，英伟达全新「阉割版」的Blackwell GPU曝光 1. 英伟达因美国出口管制在中国AI芯片市场份额从95%暴跌至50%，被国产芯片抢占市场； 2. 为应对困局推出新款阉割版Blackwell GPU，售价6500-8000美元，远低于H20的1-1.2万 美元； 3. 新芯片采用GDDR7内存技术，内存带宽约1.7TB/秒，以符合出口管制限制要求。 https://mp.weixin.qq.com/s/62VnkP-TrmhSd18CmDLWBA 二、 Claude 4如何思考？资深研究员回应，RLVR已得到验证 1. Claude 4采用可验证奖励强化学习(RLVR)范式，在编程和数学等有清晰反馈信号的领域取 得突破； 2. 当前AI Agent发展受限于高可靠性不足，但预计明年将出现能独立完成实际工作的软件工 程Agent； 3. 研究员预测到2026年底，AI将具备足够的"自我意识"，能执行复杂任务并判断自身能力边 界。 https://mp.weixin.qq.com/s/0mQ9xEKdGiSMsFqyXMJVgg https://mp.weixin.qq.com/ ...

Core Viewpoint - The article discusses a breakthrough in potential treatment for dry age-related macular degeneration (dAMD) discovered by an AI-driven research team, Future House, which utilized a multi-agent system to identify the ROCK inhibitor Ripasudil as a promising candidate for treatment [3][4][10]. Group 1: AI-Driven Research Process - The entire research process was primarily driven by AI, with human researchers only conducting laboratory experiments and writing the final paper [2][7]. - The multi-agent system named "Robin" integrated three agents (Crow, Falcon, Finch) to automate key steps in scientific discovery, including hypothesis generation, experimental design, and data analysis [18][19]. - The research was completed in approximately 10 weeks, significantly faster than traditional methods [9]. Group 2: Discovery and Validation - Robin identified Ripasudil, an existing drug approved for glaucoma treatment in Japan, as having potential effects on dAMD [4][50]. - The team consulted multiple experts in the field, who recognized the innovation and value of this discovery [5]. - Initial experiments showed that Ripasudil could enhance RPE cell phagocytic activity by 7.5 times compared to the control group, indicating its potential effectiveness [50]. Group 3: Mechanistic Insights - The research revealed that the ROCK inhibitor Y-27632 could restore RPE cell phagocytic efficiency, supporting the hypothesis generated by Robin [31]. - The analysis indicated that the lipid efflux pump ABCA1 was upregulated threefold in Y-27632 treated cells, which is significant for maintaining RPE cell health [45][47]. - The findings suggest that AI-driven research can not only identify effective therapeutic compounds but also uncover new molecular targets in disease pathways [48]. Group 4: Future Directions - Future House plans to open-source the code and data from this research next week, promoting transparency and collaboration in scientific research [12]. - The team emphasizes that further studies, including human trials, are necessary before clinical application can be considered [51].