![[fbd81d79-a59d-44d8-85bb-93a25217d60a.mp3]]

谷歌Veo3视频生成模型全球上线，限付费用户每日三段体验

简报：

• 谷歌正式在全球159个国家推出Veo3视频生成模型，为AI Pro与Ultra会员提供基于文本和图像的动态短视频创作服务。
• Veo3支持生成最长八秒、最高1080p甚至4K分辨率的短视频，具备音画同步、环境音效和多镜头叙事等多模态能力，适用于社交媒体、广告和教育等场景。
• 当前体验仅限谷歌AI Pro（每月约20美元，日限3段）及Ultra会员，所有生成视频均带可见和数字水印，未来即将支持照片生成视频功能。

相关链接：

• 谷歌在全球范围内开放 Veo 3 视频生成模型
• 谷歌Veo 3将推照片生成视频功能，Ultra版开放

Google Veo 3 Video Generation Model Launched Globally, Limited to Paid Users for Three Daily Generations 🚀

Brief:

• Google has officially rolled out the Veo 3 video generation model across 159 countries worldwide, providing AI Pro and Ultra members with services for creating dynamic short videos from text and images.
• Veo 3 supports generating short videos up to eight seconds long, with resolutions up to 1080p or even 4K. It features multimodal capabilities like audio-visual synchronization, ambient sound effects, and multi-shot storytelling, making it ideal for social media, advertising, and education. ✨
• Currently, access is limited to Google AI Pro (approximately $20/month, 3 generations/day) and Ultra members. All generated videos will include visible and digital watermarks. A feature to generate videos from photos will be supported soon. 📸

Related Links:

• 谷歌在全球范围内开放 Veo 3 视频生成模型
• 谷歌Veo 3将推照片生成视频功能，Ultra版开放

新AI模型Centaur能精确预测并模拟人类多领域决策行为

简报：

• 谷歌 DeepMind、普林斯顿大学等联合发布AI模型Centaur，首次实现在大规模认知任务中对人类行为的精准预测。
• Centaur 基于70亿参数的Llama 3.1大模型，通过Psych-101数据集训练，涵盖6万余名参与者、160个实验、超1000万个决策。
• 该模型不仅精准复现人类在感知、逻辑推理、道德判断等领域的决策，还表现出与人脑神经活动极高的相似性，甚至优于GPT-4等主流大模型。
• 研究中，模型还帮助科学家发现并证实了一种全新的人类决策策略。
• Centaur模型与Psych-101数据集均已全部开放，并计划未来拓展至儿童心理和精神病学等领域。

相关链接：

• IT之家报道
• 原论文（nature.com）

New AI Model Centaur Accurately Predicts and Simulates Human Decision-Making Across Multiple Domains

Briefing:

• Google DeepMind, Princeton University, and others have jointly released the AI model Centaur, achieving the first accurate prediction of human behavior in large-scale cognitive tasks. 🤖
• Centaur is based on the 70-billion-parameter Llama 3.1 large model, trained on the Psych-101 dataset, which includes over 60,000 participants, 160 experiments, and more than 10 million decisions.
• The model not only accurately replicates human decisions in areas such as perception, logical reasoning, and moral judgment but also demonstrates a very high similarity to human brain neural activity, even outperforming mainstream large models like GPT-4.
• During the research, the model also helped scientists discover and confirm a completely new human decision-making strategy. 🧠
• Both the Centaur model and the Psych-101 dataset have been fully open-sourced, with future plans to expand into areas such as child psychology and psychiatry. 💡

Related Links:

• IT Home Report
• Original Paper (nature.com)

美国解除西门子EDA对华出口限制，中国芯片设计获短期缓解

简报：

• 德国西门子确认，已收到美方通知，取消其对中国EDA（电子设计自动化）软体与技术的出口限制，恢复中国客户全面存取。
• 此次政策调整被认为与2025年6月中美稀土贸易协议相关，有助缓解中国芯片设计业的短期研发压力。
• 专家指出，这一解禁为战术性调整，核心光刻机等关键设备限制仍严峻，美方整体对华高科技围堵未松动。
• 行业关注此次解禁是否为临时让步，或将成为未来中美谈判筹码，相关后续政策动向仍不确定。

相关链接：

• 西门子宣布美撤除对中限令
• 中国晶片业暂解封背后　西门子官方证实EDA限制解除

US Lifts Siemens EDA Export Restrictions to China, Providing Short-Term Relief for Chip Design 🚀

Brief:

• German Siemens confirmed it has received notification from the US, lifting its export restrictions on EDA (Electronic Design Automation) software and technology to China, restoring full access for Chinese customers.
• This policy adjustment is believed to be related to the US-China rare earth trade agreement in June 2025, helping to alleviate short-term R&D pressure on China's chip design industry.
• Experts point out that this lifting of restrictions is a tactical adjustment. Restrictions on core lithography machines and other key equipment remain severe, and the US's overall high-tech containment of China has not loosened. 🚧
• The industry is watching whether this lifting of restrictions is a temporary concession or will become a bargaining chip in future US-China negotiations. Relevant subsequent policy developments remain uncertain. 🤔

Related Links:

• Siemens Announces US Lifts Restrictions on China
• Behind China's Chip Industry's Temporary Unblocking: Siemens Officially Confirms EDA Restrictions Lifted

台湾半导体产业崛起，全球布局与技术创新领跑世界

简报：

• 台美断交后，台湾半导体产业在“造山者”群体奋力推动下，从零起步打造出世界领先的完整半导体生态系，成为全球产业链核心。
• 纪录片《造山者》记录台积电与崇越等关键企业的合作和应变能力，揭示台湾以专业与高良率技术击败对手，而非偷窃技术。
• 当前地缘政治压力下，台湾企业加快全球布局，维持材料供给安全，并加速引入AI、数字化与供应链透明化管理以强化韧性。
• 政府推动“全球半导体民主供应链伙伴倡议”，促进多元、安全的国际产业合作，巩固台湾在全球供应链重组中的战略地位。
• 业界领袖指出，产生式AI投资激增、供应链短链化与ESG永续发展成为产业未来竞争关键，确保台湾持续领先全球科技创新浪潮。

相关链接：

• 挺過台美斷交！造山者們拚出台灣半導體奇蹟　專訪崇越：沒偷美國晶片
• 台灣半導體開大絕！童子賢、國科會都來開示未來走向

Taiwan's Semiconductor Industry Rises, Global Layout and Technological Innovation Lead the World

Briefing:

• After the severing of diplomatic ties between Taiwan and the US, Taiwan's semiconductor industry, driven by the "Mountain Makers," started from scratch to build a world-leading, complete semiconductor ecosystem, becoming the core of the global supply chain.
• The documentary "Mountain Makers" chronicles the collaboration and adaptability of key companies like TSMC and Topco, revealing how Taiwan defeated competitors with professionalism and high-yield technologies, rather than by stealing technology. 🎥
• Under current geopolitical pressures, Taiwanese companies are accelerating their global expansion, ensuring material supply security, and rapidly adopting AI, digitalization, and supply chain transparency management to enhance resilience.
• The government is promoting the "Global Semiconductor Democratic Supply Chain Partnership Initiative" to foster diverse and secure international industrial cooperation, solidifying Taiwan's strategic position in the global supply chain restructuring. 🤝
• Industry leaders point out that surging investments in generative AI, supply chain localization (shortening supply chains), and ESG sustainable development are becoming key factors for future industry competitiveness, ensuring Taiwan's continued leadership in the global wave of technological innovation. ✨

Related Links:

• 挺過台美斷交！造山者們拚出台灣半導體奇蹟　專訪崇越：沒偷美國晶片
• 台灣半導體開大絕！童子賢、國科會都來開示未來走向

Meta测试AI聊天机器人主动消息功能，提升用户互动与平台价值

简报：

• Meta正在测试一种新型AI聊天机器人，使其能在未经用户主动发起对话的情况下，通过Messenger和WhatsApp等平台主动向用户发送消息。
• 该功能基于Meta与数据公司Alignerr合作开发，用户可在AI Studio平台创建并定制具备记忆与个性的聊天机器人。
• 机器人只能在用户14天内发送至少5条消息后才会主动联系，且仅能发送一次消息，无回应则对话结束。
• Meta表示此举旨在增加用户参与度及平台互动，同时预计生成式AI产品到2025年将带来20亿至30亿美元收入。
• 此项功能仍处于早期测试阶段，用户隐私与安全规范受到严格审查，Meta强调机器人不能替代法律、医疗等专业服务。

相关链接：

• Meta测试AI聊天机器人主动消息功能，目标为提升用户粘性
• Meta推出主动聊天机器人，首度允许AI“先开口”

Meta Testing Proactive Messaging Feature for AI Chatbots to Boost User Engagement and Platform Value

Brief:

• Meta is testing a new type of AI chatbot that can proactively send messages to users on platforms like Messenger and WhatsApp, without the user initiating the conversation. 🤖
• This feature is developed in collaboration with data company Alignerr, allowing users to create and customize chatbots with memory and personality on the AI Studio platform.
• The chatbot will only proactively contact users after they have sent at least 5 messages within a 14-day period. It can only send one proactive message, and if there's no response, the conversation ends.
• Meta states that this initiative aims to increase user engagement and platform interaction, expecting generative AI products to generate $2-3 billion in revenue by 2025. 💰
• This feature is still in its early testing phase, with user privacy and safety protocols undergoing strict scrutiny. Meta emphasizes that chatbots cannot replace professional services like legal or medical advice. 🔒

Related Links:

• Meta Testing AI Chatbot Proactive Messaging Feature, Aiming to Boost User Stickiness
• Meta Launches Proactive Chatbots, First Time Allowing AI to "Speak First"

MiniMax推出全球首个开源大规模混合架构推理模型M1，刷新行业技术标杆

简报：

• 上海AI公司MiniMax近日正式发布全球首个开源大规模混合架构推理模型MiniMax-M1，并迅速登上权威评测榜单，位列全球开源模型第二，仅次于DeepSeek-R1-0528。
• MiniMax-M1在长文本处理与工具调用方面表现突出，支持高达100万token的上下文输入，处理能力是同类模型八倍，接近谷歌Gemini2.5Pro水平。
• MiniMax在强化学习阶段总投入仅为53.5万美元，低于业内千万级研发成本，并连续四天推出多款创新AI产品。
• MiniMax自2022年底起便布局通用人工智能，摒弃稠密架构，坚持基础模型研发，在全球大模型企业普遍收缩背景下实现技术突破，引发行业广泛关注。

相关链接：

• MiniMax发布全球首个开源大规模混合架构推理模型
• MiniMax-M1刷新全球开源大模型评测榜单
• Bloomberg报道MiniMax-M1模型发布

MiniMax Launches World's First Open-Source Large-Scale Hybrid Architecture Inference Model M1, Setting a New Industry Technical Benchmark 🚀

Brief:

• Shanghai-based AI company MiniMax recently officially released the world's first open-source large-scale hybrid architecture inference model, MiniMax-M1. It quickly climbed authoritative evaluation rankings, securing the second position globally among open-source models, trailing only DeepSeek-R1-0528. ✨
• MiniMax-M1 excels in long-text processing and tool calling, supporting up to 1 million tokens of contextual input. Its processing capability is eight times that of similar models, approaching the level of Google's Gemini 2.5 Pro.
• MiniMax's total investment in the reinforcement learning phase was only US$535,000, significantly lower than the industry's typical tens of millions in R&D costs. Furthermore, it launched multiple innovative AI products for four consecutive days.
• Since late 2022, MiniMax has been strategically developing general artificial intelligence, eschewing dense architectures and committing to foundational model R&D. Its technological breakthrough, achieved amid a general contraction among global large model companies, has garnered significant industry attention. 💡

Related Links:

• MiniMax Launches World's First Open-Source Large-Scale Hybrid Architecture Inference Model
• MiniMax-M1 Reshapes Global Open-Source Large Model Evaluation Rankings
• Bloomberg Reports on MiniMax-M1 Model Release