![[8c3c96a8-1e8e-43db-9cb5-1e2d3ae3e9ed.mp3]]

字节跳动Seedance 1.0在AI视频生成领域实现技术突破，超越谷歌Veo 3

简报：

• 字节跳动发布的AI视频生成模型Seedance 1.0，在独立评测中表现超过谷歌最新推出的Veo 3，成为视频生成技术的新标杆。
• Seedance 1.0通过空间与时间层解耦和多模态位置编码，支持复杂场景切换与多镜头叙事，在生成速度与视觉一致性方面均优于同类模型。
• 该模型在与电影导演合作开发的SeedVideoBench测试中，在遵循提示和动作真实感等多个维度取得更高分数，并能在41.4秒内生成五秒1080p视频，效率领先于Sora、Runway Gen-4与Veo 3。
• Seedance 1.0预计将于2025年6月集成至Doubao和Jimeng等平台，用于提升专业与常规创作流程。
• 尽管Veo 3创新地实现了视频与环境音效及对话的同步生成，Seedance 1.0在视觉保真度、运动稳定性及叙事连贯性等方面更具优势。

相关链接：

• Seedance1.0视频大模型首公开，性能超越Google Veo3
• 字节跳动Seedance 1.0技术细节与测试结果公布
• Seedance 1.0加速AI视频生成创新步伐

ByteDance Seedance 1.0 Achieves Breakthrough in AI Video Generation, Surpassing Google Veo 3

Briefing:

• ByteDance's AI video generation model, Seedance 1.0, has outperformed Google's latest Veo 3 in independent evaluations, setting a new benchmark for video generation technology. 🚀
• Seedance 1.0 leverages spatial and temporal layer decoupling along with multi-modal positional encoding to support complex scene transitions and multi-shot narratives. It surpasses similar models in both generation speed and visual consistency.
• In the SeedVideoBench test, developed in collaboration with film directors, the model achieved higher scores across multiple dimensions, including prompt adherence and motion realism. It can generate a five-second 1080p video in just 41.4 seconds, demonstrating superior efficiency compared to Sora, Runway Gen-4, and Veo 3. ⏱️
• Seedance 1.0 is expected to be integrated into platforms like Doubao and Jimeng by June 2025, aiming to enhance both professional and conventional creative workflows.
• While Veo 3 innovatively achieves synchronized generation of video with environmental sound effects and dialogue, Seedance 1.0 holds a stronger advantage in visual fidelity, motion stability, and narrative coherence. ✨

Related Links:

• Seedance1.0视频大模型首公开，性能超越Google Veo3
• 字节跳动Seedance 1.0技术细节与测试结果公布
• Seedance 1.0加速AI视频生成创新步伐

苹果Speech API视频转录速度创新纪录，远超同类AI工具

简报：

• 苹果在2025年WWDC推出的全新Speech API，实现了将34分钟7GB的4K视频在45秒内完成转录。
• 这一速度比OpenAI的Whisper（MacWhisper V3Turbo版本）快55%，后者需101秒完成同样任务。
• 测试显示，Speech API基础的Yap工具转录效率也远高于VidCap和MacWhisper V2等主流工具。
• 新Speech技术为内容创作者和教育工作者带来显著便利，提升视频内容处理效率。
• 尽管各工具在专有名词识别方面略有误差，Yap凭借本地化运算能力在速度上表现突出。

相关链接：

• aibase.com新闻报道

Apple Speech API Sets New Video Transcription Speed Record, Greatly Exceeding Other AI Tools

Brief:

• Apple's new Speech API, introduced at WWDC 2025, transcribed a 34-minute, 7GB 4K video in just 45 seconds. ⚡
• This speed is 55% faster than OpenAI's Whisper (MacWhisper V3Turbo version), which took 101 seconds to complete the same task.
• Tests show that Yap, a tool built on the Speech API, also boasts significantly higher transcription efficiency than mainstream tools like VidCap and MacWhisper V2.
• The new Speech technology brings significant convenience to content creators and educators, greatly improving video content processing efficiency. ✨
• Although various tools showed slight errors in proper noun recognition, Yap stood out in speed due to its localized computing capabilities. 👍

Related Links:

• aibase.com News Report

OpenAI宣布GPT-4.5将于7月中旬从API移除，开发者面临转型压力

简报：

• OpenAI已通知开发者，GPT-4.5预览版将于2025年7月14日从官方API下线，尽管该模型仍可供ChatGPT个人用户使用。
• 许多依赖GPT-4.5的开发者对这一决定表示迷惑与失落，不得不抓紧时间调整产品和流程，迁移至GPT-4.1或其他模型。
• OpenAI早在今年4月公布了淘汰计划，并建议开发者通过其在线论坛获取迁移帮助；有分析认为此次下线或与高计算成本有关。

相关链接：

• OpenAI moves forward with GPT-4.5 deprecation in API, triggering developer anguish and confusion
• OpenAI plans to phase out GPT-4.5, its largest-ever AI model, from its API
• AI Base 中文报道
• IT之家新闻

OpenAI Announces GPT-4.5 API Removal by Mid-July, Developers Face Transition Pressure

Brief:

• OpenAI has notified developers that the GPT-4.5 preview version will be deprecated from the official API on July 14, 2025, though the model will remain available for individual ChatGPT users. 🤖
• Many developers relying on GPT-4.5 have expressed confusion and disappointment over this decision, rushing to adjust products and processes and migrate to GPT-4.1 or other models. ⚙️
• OpenAI announced the deprecation plan back in April this year, advising developers to seek migration assistance via its online forums; some analysts believe the removal may be linked to high computational costs. 💸

Related Links:

• OpenAI moves forward with GPT-4.5 deprecation in API, triggering developer anguish and confusion
• OpenAI plans to phase out GPT-4.5, its largest-ever AI model, from its API
• AI Base 中文报道
• IT之家新闻

华为汪涛预测2030年全球AI机器人数量将达30亿

简报：

• 在2025 MWC上海期间，华为常务董事汪涛表示，预计到2030年，全球将有近30亿AI机器人进入工厂和家庭。
• 汪涛指出，5G-A网络的快速发展和AI技术的融合正推动个人通信、行业物联和移动网络的革新，AI应用月活跃用户数已突破12亿。
• 他还预测到2030年，每年将新增超8000万辆网联汽车，AI机器人将在工业和家庭应用中快速普及。

相关链接：

• 华为汪涛：预计2030年全球将有近30亿AI机器人进入工厂、家庭

Huawei's David Wang Predicts Global AI Robot Population to Reach 3 Billion by 2030 🤖🌍

Briefing:

• During MWC Shanghai 2025, David Wang, Huawei's Executive Director, stated that nearly 3 billion AI robots are expected to enter factories and homes globally by 2030. 🏠
• Wang Tao pointed out that the rapid development of 5G-A networks and the integration of AI technology are driving innovation in personal communications, industrial IoT, and mobile networks. Monthly active users of AI applications have already exceeded 1.2 billion.
• He also predicted that by 2030, over 80 million connected vehicles will be added annually, and AI robots will rapidly become widespread in industrial and home applications.

Related Link:

• Huawei David Wang: Nearly 3 Billion AI Robots Expected in Factories and Homes Globally by 2030

华为申报四芯片封装专利，或助力突破美国AI芯片禁令

简报：

• 华为近日申报“四芯片封装”专利，可能用于下一代AI芯片Ascend 910D，其设计被指与英伟达最新Rubin Ultra架构相似；
• 新技术采用类似桥接架构，可结合多组高频宽内存（HBM）提升AI训练效能，外媒分析认为此举或使华为在封装领域逐步赶超台积电，并有助于绕过美国出口限制；
• 华为被曝早前囤积大量AI芯片裸晶，通过多裸晶封装方案，提升自主高性能AI芯片产能，目标缩小与英伟达在AI GPU性能上的差距；
• 台湾已将华为及其合作方中芯国际列入出口实体管理黑名单，但业内人士认为华为通过“先囤料、再封装”等方式，仍有能力保持一定AI芯片产能。

相关链接：

• 华为新专利曝光挑战台积电英伟达
• 华为囤积200万裸晶挑战NVIDIA

Huawei Files Four-Chip Packaging Patent, Potentially Aiding Breakthrough Against US AI Chip Ban

Brief:

• Huawei recently filed for a "four-chip packaging" patent, potentially for use in its next-generation AI chip, the Ascend 910D. Its design is reportedly similar to NVIDIA's latest Rubin Ultra architecture. 🚀
• The new technology adopts a bridge-like architecture capable of combining multiple High Bandwidth Memory (HBM) stacks to enhance AI training performance. Foreign media analysis suggests this move could gradually enable Huawei to catch up with TSMC in packaging technology and help circumvent US export restrictions.
• Huawei is reported to have previously stockpiled a large quantity of AI chip bare dies. Through a multi-die packaging solution, it aims to boost its independent high-performance AI chip production capacity, with the goal of narrowing the performance gap with NVIDIA's AI GPUs. 💪
• Taiwan has blacklisted Huawei and its partner SMIC under its export entity management list, but industry insiders believe Huawei, through methods like "stockpiling materials first, then packaging," can still maintain a certain level of AI chip production capacity. 💡

Related Links:

• Huawei's new patent exposed, challenging TSMC and NVIDIA
• Huawei stockpiles 2 million bare dies, challenging NVIDIA

CMU与英伟达联合发布Multiverse模型推动大语言模型并行生成提速

简报：

• 卡耐基梅隆大学（CMU）与英伟达联合发布新型生成模型Multiverse，实现大语言模型（LLM）推理过程的并行化，加快生成速度。
• Multiverse采用类似MapReduce的结构，将生成任务分解为多个可并行执行的子任务，并无损合并结果，大幅提高硬件资源利用率。
• 实验显示，Multiverse-32B模型在同等上下文长度下性能较传统自回归方式提升近2%，并能根据需求灵活切换顺序与并行生成模式。
• 团队已开源Multiverse生态系统，包括数据、模型权重和训练细节，推动相关领域的进一步研究和应用拓展。

相关链接：

• CMU与英伟达携手推出Multiverse，实现超高速并行生成

CMU and NVIDIA Jointly Launch Multiverse Model to Accelerate Large Language Model Parallel Generation

Brief:

• Carnegie Mellon University (CMU) and NVIDIA have jointly released a new generative model, Multiverse, enabling parallelization of Large Language Model (LLM) inference processes to accelerate generation speed. 🚀
• Multiverse adopts a MapReduce-like structure, breaking down generation tasks into multiple parallelizable subtasks and merging results without loss, significantly improving hardware resource utilization. ✨
• Experiments show that the Multiverse-32B model improves performance by nearly 2% compared to traditional autoregressive methods at equivalent context lengths, and can flexibly switch between sequential and parallel generation modes as needed.
• The team has open-sourced the Multiverse ecosystem, including data, model weights, and training details, to promote further research and application expansion in related fields. 💻

Related Links:

• CMU and NVIDIA Partner to Launch Multiverse for Ultra-Fast Parallel Generation