单词学习：![[ba54e45a-8ba5-47ca-91b4-5e1845d4becc.mp3]]

新闻播报：![[9ec1ad85-b248-41c7-9280-561b2fd004f7.mp3]]

OpenAI拟斥巨资扩充服务器以缓解算力瓶颈

简报：

• 据The Information报道，OpenAI计划在未来五年内追加1000亿美元投资，用于租用备用服务器，以应对因计算能力严重受限导致的产品发布推迟问题。
• 增加服务器投资旨在应对用户需求的突发增长，并为未来模型的训练提供支持。
• 预计到2030年，OpenAI在服务器租赁上的总支出将达到3500亿美元，预计在2029年前的现金流出量将达到1150亿美元。

相关链接：

• OpenAI 计划未来五年追加 1000 亿美元服务器投资

OpenAI Plans Massive Server Expansion to Ease Computing Bottlenecks 🚀

Brief:

• According to The Information, OpenAI plans to invest an additional $100 billion over the next five years to lease backup servers 💡, addressing product launch delays caused by severe computing power limitations.
• This increased server investment aims to cope with sudden surges in user demand and support the training of future models 📈.
• By 2030, OpenAI's total expenditure on server leasing is projected to reach $350 billion, with an estimated cash outflow of$115 billion by 2029.

Related Links:

• OpenAI Plans to Invest an Additional $100 Billion in Servers Over the Next Five Years

字节跳动发布豆包翻译大模型，主打低价与高性能

简报：

• 字节跳动旗下火山引擎近日发布了全新的通用翻译大模型“豆包翻译模型”（Doubao-Seed-Translation）。
• 该模型支持多达28种语言的互译能力，官方宣称其翻译性能已达到或超越GPT-4o和Gemini-2.5-Pro。
• 模型定价具有竞争力，每百万字符输入收费1.20元，输出收费3.60元。

相关链接：

• 火山引擎官方定价文档

ByteDance Launches Doubao Translation Large Model, Featuring Low Cost and High Performance

Briefing:

• ByteDance's Volcengine recently launched a brand-new general-purpose translation large model, the "Doubao Translation Model" (Doubao-Seed-Translation). 🎉
• The model supports bidirectional translation for up to 28 languages, with official claims stating its translation performance has met or surpassed GPT-4o and Gemini-2.5-Pro. 🚀
• The model's pricing is highly competitive, costing 1.20 CNY per million input characters and 3.60 CNY per million output characters. 💰

Related Links:

• Volcengine Official Pricing Documentation

OpenAI将推出付费专享的高算力新功能

简报：

• OpenAI首席执行官Sam Altman宣布，公司将在未来几周内推出一系列需要更多计算资源的新服务。
• Altman表示，由于相关成本高昂，这些新功能初期将仅对Pro订阅用户开放。
• 此外，部分新产品可能还将收取额外费用。
• OpenAI的长期目标是尽可能降低智能服务的成本，使其更广泛地普及。

相关链接：

• OpenAI 计划未来五年追加 1000 亿美元服务器投资
• OpenAI CEO透露将推计算密集型新功能，部分仅限Pro用户

OpenAI to Launch New High-Compute Features for Paid Subscribers 🚀

Briefing:

• OpenAI CEO Sam Altman announced that the company will introduce a range of new services requiring more computational resources in the coming weeks.
• Altman stated that due to high associated costs, these new features will initially be exclusive to Pro subscribers.
• Furthermore, some new products might incur additional charges. 💰
• OpenAI's long-term goal is to reduce the cost of intelligent services as much as possible to make them more widely accessible. 🌐

Related Links:

• OpenAI Plans to Invest an Additional $100 Billion in Servers Over the Next Five Years
• OpenAI CEO Reveals Plans for Compute-Intensive New Features, Some Exclusive to Pro Users

曝Meta AR眼镜路线图：2027年推双屏及真AR眼镜

简报：

• 据彭博社记者马克・古尔曼透露，Meta计划于2027年推出一款配备双显示屏的智能眼镜，并在同年晚些时候发布一款真正的全息AR眼镜。
• 此消息在Meta最新款Ray-Ban Display智能眼镜发布后传出，该产品在近期的发布会演示中遭遇了AI指令失效、视频通话无法接通等技术故障。
• 古尔曼称，目前Meta的Ray-Ban Display智能眼镜正在亏本销售，以巩固其在未来计算领域的地位。

相关链接：

• 古尔مان：Meta 将于 2027 年推出双屏智能眼镜，以及真 AR 眼镜
• AI眼鏡發表會翻車！Meta認了　技術長急認錯曝主因

Meta AR Glasses Roadmap Leaked: Dual-Screen & True AR Glasses Slated for 2027

Brief:

• According to Bloomberg journalist Mark Gurman, Meta plans to launch a pair of smart glasses equipped with dual displays in 2027, followed by a true holographic AR glasses later the same year. 👓
• This news comes after the release of Meta's latest Ray-Ban Display smart glasses, which faced technical glitches during recent launch event demonstrations, including AI command failures and inability to connect video calls. 😬
• Gurman states that Meta's Ray-Ban Display smart glasses are currently being sold at a loss to solidify its position in the future of computing. 📉

Related Links:

• Gurman: Meta to Launch Dual-Screen Smart Glasses and True AR Glasses in 2027
• AI Glasses Launch Flops! Meta Admits Fault, CTO Quickly Apologizes and Reveals Main Reason

联发科发布天玑9500，主打端侧AI与游戏性能

简报：

• 联发科技于9月22日正式发布旗舰芯片天玑9500，该芯片采用台积电第三代3纳米制程工艺。
• 芯片以“5G Agentic AI”为核心，整合全大核CPU、旗舰级GPU与NPU 990，主打端侧最强AI算力与“3A级冷劲游戏体验”。
• 性能方面，相较上一代，其单核CPU性能提升32%，多核提升17%，GPU性能提升33%，AI算力跃升超过一倍。
• 影像与功能上，首次支持设备端4K画质图片生成、4K 120FPS防震录影，并提供AI选网与AI定位等智慧通讯功能。
• 首批搭载天玑9500的智能手机预计将于2025年第四季度问世。

相关链接：

• 天玑9500正式亮相 联发科旗舰芯片火力全开震撼登场

MediaTek Unveils Dimensity 9500, Highlighting On-Device AI and Gaming Performance

Brief:

• MediaTek officially launched its flagship chip, the Dimensity 9500, on September 22. The chip utilizes TSMC's third-generation 3nm process technology.
• Centered around "5G Agentic AI," the Dimensity 9500 integrates a full large-core CPU, flagship-grade GPU, and NPU 990. It aims to deliver the strongest on-device AI computing power and a "AAA-level cool and powerful gaming experience." 🎮
• In terms of performance, compared to its predecessor, its single-core CPU performance is boosted by 32%, multi-core by 17%, and GPU performance by 33%. Remarkably, its AI computing power more than doubles. 🚀
• For imaging and features, it introduces support for on-device 4K image generation, 4K 120FPS anti-shake video recording, and smart communication functions such as AI network selection and AI positioning.
• The first smartphones powered by the Dimensity 9500 are expected to hit the market in Q4 2025. ✨

Related Links:

• Dimensity 9500 Officially Debuts: MediaTek's Flagship Chip Unleashes Full Power

中科院发布类脑脉冲大模型，长文本处理速度提升百倍

简报：

• 中国科学院自动化研究所发布全球首款大规模类脑脉冲大模型SpikingBrain，处理400万token超长文本时速度比主流Transformer模型快100多倍，且仅需2%的训练数据。
• 该模型采用“混合线性注意力架构”，将长文本处理的计算复杂度从二次方级别降至线性级别。
• 模型通过引入“自适应阈值脉冲神经元”模拟人脑工作方式，以事件驱动机制节省能耗，计算稀疏度达69.15%。
• SpikingBrain包含70亿和760亿参数两个版本，相关技术与代码已在GitHub及魔搭平台开源。

相关链接：

• SpikingBrain-7B GitHub 开源地址

CAS Releases Brain-Inspired Spiking Large Model, Long Text Processing Speed Increased 100x 🧠

Briefing:

• The Institute of Automation, Chinese Academy of Sciences (CASIA), has released SpikingBrain, the world's first large-scale brain-inspired spiking large model. It processes ultra-long texts of 4 million tokens over 100 times faster than mainstream Transformer models, and requires only 2% of the training data. 🚀
• The model employs a "hybrid linear attention architecture," reducing the computational complexity of long text processing from quadratic to linear.
• By introducing "adaptive threshold spiking neurons," the model simulates how the human brain works, saving energy through an event-driven mechanism, achieving a computational sparsity of 69.15%.
• SpikingBrain is available in two versions with 7 billion and 76 billion parameters. Related technologies and code have been open-sourced on GitHub and the ModelScope platform. 🌐

Related Links:

• SpikingBrain-7B GitHub Open-Source Address