新闻播报：![[54c9b861-6e95-4b87-b880-f98a3e64d67f.mp3]]

谷歌发布“可学习推理记忆”框架，AI智能体能从错误中自我进化

简报：

• 谷歌最新研究提出名为“Reasoning Memory”（可学习的推理记忆）的框架，旨在解决AI智能体（Agent）无法从经验中学习、重复犯错的核心痛点。
• 该框架通过让AI系统地记录、概括并重用自身的推理过程与结果，将具体经验转化为通用规则，形成可学习的“记忆”。
• 在新任务中，AI智能体可以调用这些记忆来优化决策，从而实现像人类一样“从错误中学习”并持续自我改进。
• 实验显示，配备该框架的智能体在复杂任务中的性能显著提升，标志着AI正从静态执行向动态成长演进。

相关链接：

• Google重磅突破！AI Agent 实现“自我进化”，从错误中学习变身超级智能体

Google Releases "Learnable Reasoning Memory" Framework, AI Agents Evolve from Mistakes 🚀

Brief:

• Google's latest research introduces a framework called "Reasoning Memory," designed to address the core problem of AI agents being unable to learn from experience and repeatedly making the same mistakes.
• This framework enables AI to systematically record, generalize, and reuse its own reasoning processes and results, transforming specific experiences into general rules, thereby forming a learnable "memory."
• In new tasks, AI agents can leverage these memories to optimize decision-making, allowing them to "learn from mistakes" and continuously self-improve, much like humans do.
• Experiments show that agents equipped with this framework exhibit significantly improved performance in complex tasks, marking AI's evolution from static execution to dynamic growth. 🌱

Related Links:

• Google's Major Breakthrough! AI Agents Achieve "Self-Evolution," Learn from Mistakes to Become Super Agents

国产AI助手跟进记忆功能，通义千问与豆包开启内测

简报：

• 阿里巴巴旗下“通义千问”与字节跳动旗下“豆包”已相继启动记忆功能的内部测试，旨在从“即时问答工具”向“长期私人助理”转型，对标OpenAI的ChatGPT。
• 通义千问的记忆功能已在小范围上线，采用“显式记忆”设计，由用户主动触发并经AI确认后存储，以增强用户的控制感。
• 豆包于10月10日更新了《隐私政策》并发布《记忆功能FAQ》，明确用户可随时开启、关闭、查看和删除记忆，且该功能不适用于图片、文件等内容。
• 此举被视为国内大模型应用从追求参数规模转向注重提升用户个性化体验的重要标志。

相关链接：

• 国内AI助手通义千问和豆包开启记忆功能，力争超越ChatGPT

Domestic AI Assistants Implement Memory Features, Tongyi Qianwen and Doubao Begin Internal Testing 🤖

Brief:

• Alibaba's Tongyi Qianwen and ByteDance's Doubao have successively initiated internal testing for memory functions 🧠, aiming to transition from 'instant Q&A tools' to 'long-term personal assistants,' benchmarking OpenAI's ChatGPT.
• Tongyi Qianwen's memory function has been partially launched to a small user base, adopting an 'explicit memory' design where memories are actively triggered by users and stored after AI confirmation, to enhance user control.
• Doubao updated its 'Privacy Policy' and released 'Memory Function FAQ' on October 10th, clarifying that users can enable, disable, view, and delete memories at any time, and that this function does not apply to content such as images or files.
• This move is regarded as a significant sign of domestic large model applications shifting from pursuing parameter scale to focusing on enhancing personalized user experience. ✨

Related Link:

• Domestic AI Assistants Tongyi Qianwen and Doubao Enable Memory Features, Striving to Surpass ChatGPT

OpenAI携手阿根廷，启动拉美首个“星际之门”AI数据中心项目

简报：

• OpenAI宣布在阿根廷建设其“星际之门”项目在拉丁美洲的首个AI基础设施——Stargate Argentina。
• 该项目将与阿根廷能源企业Sur Energy合作，共同建设一座能源规模达500MW的大型AI数据中心。
• 项目投资总额最高可达250亿美元，将在阿根廷的RIGI税收减免计划框架下构建。

相关链接：

• OpenAI“星际之门”拉美首站：将在阿根廷建设 500MW 规模 AI 数据中心
• OpenAI to build Stargate in Argentina

OpenAI Teams Up with Argentina to Launch Latin America's First "Stargate" AI Data Center Project 🚀

Brief:

• OpenAI has announced the construction of its "Stargate" project's first AI infrastructure in Latin America – Stargate Argentina – in Argentina.
• The project will involve a collaboration with Argentine energy company Sur Energy to jointly build a large-scale AI data center with an energy capacity of 500MW. 💡
• The total project investment could reach up to $25 billion, to be developed under Argentina's RIGI tax incentive program. 💰

Related Links:

• OpenAI“星际之门”拉美首站：将在阿根廷建设 500MW 规模 AI 数据中心
• OpenAI to build Stargate in Argentina

Sora安卓版上线在即，北美率先开启预注册

简报：

• OpenAI的AI视频生成应用Sora即将登陆安卓平台，应用已在Google Play商店中出现。
• 目前该应用尚无法下载使用，但已面向美国和加拿大的用户开启预注册。
• 该发布策略与Sora在iOS平台的初期策略相似，后者首发时也仅限于北美地区。
• Sora在iOS平台表现亮眼，发布后曾于数日内突破100万次下载。

相关链接：

• Sora by OpenAI 即将登陆安卓平台，预注册已在北美开启

Sora Android Version Imminent, Pre-registration Kicks Off in North America

Briefing:

• OpenAI's AI video generation app, Sora, is set to launch on Android, with the application already appearing in the Google Play Store. 🚀
• While the app is not yet available for download and use, pre-registration has opened for users in the United States and Canada. 🇨🇦🇺🇸
• This release strategy mirrors Sora's initial approach on the iOS platform, which also launched exclusively in North America.
• Sora achieved remarkable success on the iOS platform, surpassing 1 million downloads within just a few days of its release.

Related links:

• Sora by OpenAI is coming to Android, pre-registration open in North America

Windows 10支持明日终止，微软提供升级或付费续期方案

简报：

• 微软将于10月14日正式终止对Windows 10的支持，全球仍有约43%的Windows用户（数亿台设备）将不再获得安全更新，面临更高的网络攻击风险。
• 微软提供了两个替代方案：一是为符合硬件条件的电脑免费升级至Windows 11；二是注册“延伸安全性更新”（ESU）计划。
• ESU计划可在未来12个月内持续提供关键安全更新，但企业用户需为每台设备支付61美元，不符合免费条件的个人用户则需支付30美元或用点数兑换。
• 消费者团体批评称，此举迫使许多仍可正常使用的电脑被淘汰，对消费者经济和环境造成压力。

相关链接：

• 43%用戶還在Win10！數億裝置失去安全更新 微軟2替代方案出爐

Windows 10 Support Ends Tomorrow, Microsoft Offers Upgrade or Paid Extension Plans

Brief:

• Microsoft will officially end support for Windows 10 on October 14th. Approximately 43% of Windows users globally (hundreds of millions of devices) will no longer receive security updates, facing a higher risk of cyberattacks. 💻
• Microsoft has provided two alternative solutions: 1. A free upgrade to Windows 11 for PCs that meet the hardware requirements; 2. Registering for the "Extended Security Update" (ESU) program.
• The ESU program will continue to provide critical security updates for the next 12 months. However, enterprise users will need to pay $61 per device, while individual users who don't qualify for the free upgrade will have to pay$30 or redeem points. 💰
• Consumer groups criticize this move, stating it forces many still-functional computers into obsolescence, putting a strain on consumer finances and the environment. ♻️

Related Link:

• 43% of users still on Win10! Hundreds of millions of devices lose security updates, Microsoft unveils 2 alternative solutions

xAI挖角英伟达专家，加速AI游戏与机器人研发

简报：

• 埃隆・马斯克的人工智能公司 xAI 为加速“世界模型”研发，已从竞争对手 NVIDIA 挖来核心研究员 Zeeshan Patel 和 Ethan He。
• xAI 计划利用该技术在2026年底前推出首款能实时反应玩家行为的3D生成游戏。
• 长期来看，该技术还将扩展至机器人和自动驾驶领域，与特斯拉等马斯克旗下公司形成协同效应。

相关链接：

• 马斯克挖角 NVIDIA 核心团队，xAI 加速 “世界模型” 研发
• Grok 即将具备检测 AI 生成视频的能力

xAI Poaches NVIDIA Experts, Accelerating AI Game and Robotics R&D 🚀

Briefing:

• Elon Musk's AI company xAI has poached key researchers Zeeshan Patel and Ethan He from competitor NVIDIA to accelerate its "world model" development.
• xAI plans to leverage this technology to launch its first 3D generative game that reacts to player behavior in real-time by the end of 2026. 🎮
• In the long run, this technology will also extend to robotics and autonomous driving, creating synergies with other Musk-owned companies like Tesla. 🤖

Related Links:

• Musk Poaches NVIDIA Core Team, xAI Accelerates "World Model" R&D
• Grok Soon to Be Capable of Detecting AI-Generated Videos

蚂蚁集团开源万亿参数大模型Ling-1T

简报：

• 蚂蚁集团宣布开源其最新的旗舰大模型 Ling-1T，该模型拥有万亿级别参数，是目前已知使用FP8低精度模式训练的最大基座模型。
• Ling-1T 在处理每个token时，仅激活约500亿参数，从而大幅降低运算成本。
• 模型训练分为预训练、中训练和后训练三个阶段，预训练数据量超过20万亿个token，并强化了推理能力。
• 在多项测试中，Ling-1T 在数学推理和代码生成能力上表现出色。
• 蚂蚁团队表示，目前模型在处理超长上下文时成本较高，正在研究新的混合注意力架构以解决该问题。

相关链接：

• 蚂蚁集团发布万亿参数模型 Ling-1T，超越 GPT-5 成新标杆

Ant Group Open-Sources Trillion-Parameter Large Model Ling-1T

Briefing:

• Ant Group announced the open-sourcing of its latest flagship large language model, Ling-1T. This model boasts a trillion-level parameter count, making it the largest base model known to be trained using FP8 low-precision mode. 🚀
• When processing each token, Ling-1T activates only approximately 50 billion parameters, significantly reducing computational costs.
• The model's training is divided into three stages: pre-training, intermediate training, and post-training. The pre-training dataset exceeds 20 trillion tokens, and its reasoning capabilities have been enhanced. 🧠
• In various tests, Ling-1T demonstrated outstanding performance in mathematical reasoning and code generation capabilities.
• The Ant team stated that the model currently incurs high costs when processing extremely long contexts and is researching new hybrid attention architectures to address this issue. ✨

Related Links:

• Ant Group Launches Trillion-Parameter Model Ling-1T, Surpassing GPT-5 as a New Benchmark