单词学习：![[1b249e33-fa48-4811-a71a-05a3113c5b0a.mp3]]

新闻播报：![[caee7ea1-6a34-4c70-ab74-eaeac1b5b594.mp3]]

OpenAI发布GPT-5.1-Codex-Max模型API，定价不变性能提升

简报：

• OpenAI已面向所有开发者开放其最强编程模型GPT-5.1-Codex-Max的API访问权限，定价与标准GPT-5模型保持一致。
• 该模型专为长时间运行任务设计，内部测试数据显示能够连续执行任务超过24小时而不掉线。
• 在性能上，Max版本在SWE-Bench Verified等测试中得分全面超越常规版，并专门针对Windows开发环境进行了优化。
• 目前，包括Cursor、GitHub Copilot在内的多款主流编程工具已完成对该模型的集成。

相关链接：

• IT之家：加量不加价：OpenAI 最强编程 AI 模型 API 开放，连续写代码可超 24 小时

OpenAI Releases GPT-5.1-Codex-Max Model API: Same Price, Enhanced Performance

Briefing:

• OpenAI has opened API access to its most powerful coding model, GPT-5.1-Codex-Max, for all developers, with pricing remaining consistent with the standard GPT-5 model. 🚀
• This model is specifically designed for long-running tasks, with internal test data showing it can continuously execute tasks for over 24 hours without interruption. 🕰️
• In terms of performance, the Max version comprehensively outperforms the regular version in tests such as SWE-Bench Verified and has been specifically optimized for Windows development environments.
• Currently, several mainstream coding tools, including Cursor and GitHub Copilot, have completed integration with this model. ✨

Related Links:

• IT之家: Same Price, More Power: OpenAI's Strongest Coding AI Model API Now Open, Can Write Code Continuously for Over 24 Hours

OpenAI全股票收购Neptune，强化大模型训练监控

简报：

• OpenAI宣布将以全股票形式收购波兰AI初创公司Neptune，交易金额不足4亿美元。
• Neptune专注于AI训练过程的跟踪与可视化，其团队和技术将被整合，以提升GPT大模型的训练监控、调试和审计能力。
• 这是OpenAI在2025年内进行的第四次收购，Neptune的对外服务计划在2026年第一季度前逐步关停。
• 此次收购旨在缩短下一代GPT模型的实验周期，并降低超大规模集群的调试成本。

OpenAI Acquires Neptune in All-Stock Deal, Enhancing Large Model Training Monitoring

Briefing:

• OpenAI announced its acquisition of Polish AI startup Neptune in an all-stock deal valued at less than $400 million. 🤝
• Neptune specializes in tracking and visualizing AI training processes. Its team and technology will be integrated to enhance the monitoring, debugging, and auditing capabilities of GPT large models. 🔬
• This marks OpenAI's fourth acquisition within 2025. Neptune's external services are planned to be gradually phased out before Q1 2026.
• The acquisition aims to shorten the experimental cycle for next-generation GPT models and reduce debugging costs for hyperscale clusters. 🚀

AMD确认AI芯片MI308将重返中国，但需支付15%销售税

简报：

• AMD公司CEO苏姿丰于12月4日亲口证实，旗下AI芯片Instinct MI308已获得美国政府许可，获准出口至中国大陆。
• 作为恢复出口的条件，AMD需就相关销售额向美国政府缴纳15%的税金，苏姿丰表示公司已“准备好缴税也准备好出货”。
• 此举解除了自2025年4月起美国商务部对MI308芯片实施的出口限制，该限制曾导致AMD预估损失高达8亿美元的营收。
• 据悉，此项“用钱买出口权”的协议是美国政府与AMD、英伟达等公司达成的折衷方案，旨在平衡技术管制与企业经济利益。

相关链接：

• 蘇姿丰親自證實MI308晶片獲准入中國市場！下一個会是輝達吗？

AMD Confirms MI308 AI Chip's Return to China, Subject to 15% Sales Tax

Brief:

• AMD CEO Lisa Su personally confirmed on December 4 that the company's AI chip, Instinct MI308, has received U.S. government approval for export to mainland China. ✅
• As a condition for resuming exports, AMD will be required to pay a 15% tax on related sales revenue to the U.S. government. Su stated that the company is "ready to pay the tax and ready to ship." 💰
• This move lifts the export restrictions imposed by the U.S. Department of Commerce on the MI308 chip since April 2023, which had previously led to AMD's estimated revenue losses of up to $800 million. 📉
• Reportedly, this "pay-to-export" agreement is a compromise reached between the U.S. government and companies like AMD and Nvidia, aiming to balance technology control with corporate economic interests.

Related links:

• Lisa Su personally confirms MI308 chip approved for China market! Will Nvidia be next?

快手可灵数字人升级，视频生成时长增至5分钟

简报：

• 快手旗下可灵AI宣布，全新一代可灵数字人2.0已全量上线。
• 用户仅需通过“上传角色图、添加配音内容、描述角色表现”三个步骤，即可生成数字人视频。
• 新版本实现三大突破：大幅提升了体态、手势、表情等表现力；实现了更精准的手部及口型控制；支持生成最长5分钟的视频内容，较此前版本的1分钟大幅延长。
• 该功能基于多模态理解与视频生成模型，采用Transformer的DiT架构，能确保口型精准同步及视频全程的角色一致性。

相关链接：

• 快手可灵数字人 2.0 全量上线：三步做到能说会演，最长 5 分钟

Kuaishou Koling AI Digital Human Upgrades, Video Generation Extends to 5 Minutes

Brief:

• Kuaishou's Koling AI announced that the new generation of Koling Digital Human 2.0 has been fully launched. 🎉
• Users can generate digital human videos in just three steps: "upload character image, add voiceover content, and describe character performance."
• The new version achieves three major breakthroughs: significantly enhanced expressiveness in posture, gestures, and facial expressions; more precise hand and lip-sync control; and support for generating video content up to 5 minutes long, a substantial increase from the previous version's 1 minute. 🚀
• This feature is based on a multi-modal understanding and video generation model, utilizing a Transformer-based DiT architecture, which ensures accurate lip-sync and consistent character performance throughout the video. ✨

Related Links:

• Kuaishou Koling Digital Human 2.0 Fully Launched: Capable of Speaking and Performing in Three Steps, Up to 5 Minutes Long

谷歌Gemini 3 Deep Think公测，推理能力达奥赛金牌水平

简报：

• 谷歌于2025年12月5日面向Gemini Ultra订阅用户推出Gemini 3 Deep Think（深度思考）模式的公开测试。
• 该模式调用更多计算资源并采用先进的并行推理技术，旨在提升处理复杂数学、科学及逻辑问题的能力。
• 其变体模型曾在国际数学奥林匹克竞赛（IMO）中达到“金牌标准”，并在Humanity’s Last Exam等多个AI基准测试中取得行业顶尖成绩。
• 此次发布被视为对竞争对手OpenAI的有力回击，后者曾声称拥有同级模型但尚未向公众开放。

相关链接：

• 截胡 OpenAI：谷歌率先公测“奥数金牌级”推理 AI 模型 Gemini 3 Deep Think

Google Gemini 3 Deep Think Public Beta Launched, Reasoning Reaches Olympiad Gold Medal Level 🧠✨

Briefing:

• Google launched the public beta of Gemini 3 Deep Think (Deep Thinking) mode on December 5, 2025, for Gemini Ultra subscribers.
• This mode utilizes increased computational resources and advanced parallel reasoning techniques, aiming to enhance its capability in solving complex mathematical, scientific, and logical problems. 🚀
• A variant of this model previously achieved "gold medal standard" in the International Mathematical Olympiad (IMO) and secured top industry scores in several AI benchmarks, including Humanity's Last Exam. 🏅
• This release is seen as a strong counter to competitor OpenAI, which had previously claimed to possess a similar-tier model but has not yet made it publicly available.

Related Links:

• Google Beats OpenAI to the Punch: Public Beta Launched for "Olympiad Gold Medal Level" Reasoning AI Model Gemini 3 Deep Think

Meta战略转向AI硬件，挖角苹果设计高管组建新团队

简报：

• Meta宣布组建一个新的创意工作室，由前苹果设计副总裁艾伦・戴伊（Alan Dye）挂帅，另一位前苹果设计领袖Billy Sorrentino也一同加入。
• 该全明星设计团队将隶属于Reality Labs，首要攻坚目标是下一代AI眼镜与可穿戴设备，旨在将智能视为设计素材并融合时尚与科技。
• 此举正值Meta将投资重心从“元宇宙”转向AI眼镜和可穿戴设备，并可能削减元宇宙部门高达30%的预算。

相关链接：

• Meta 组建全明星设计团队：前苹果副总裁 Alan Dye 挂帅，攻坚 AI 眼镜
• Meta 或将在明年削减高达30% 的 “元宇宙” 预算

Meta's Strategic Shift to AI Hardware, Poaching Apple Design Executives to Form New Team

Brief:

• Meta has announced the formation of a new creative studio, led by former Apple VP of Design, Alan Dye. Another former Apple design leader, Billy Sorrentino, has also joined the team. 🚀
• This all-star design team will be part of Reality Labs, with their primary focus on developing next-generation AI glasses and wearable devices. The goal is to treat intelligence as a design material, blending fashion and technology seamlessly. ✨
• This move comes as Meta shifts its investment focus from the "metaverse" to AI glasses and wearables, potentially cutting the metaverse division's budget by up to 30%. 📉

Related Links:

• Meta Forms All-Star Design Team: Former Apple VP Alan Dye to Lead, Tackling AI Glasses
• Meta May Cut "Metaverse" Budget by Up to 30% Next Year

微软发布轻量级实时语音模型，支持90分钟超长音频生成

简报：

• 微软开源了一款名为VibeVoice-Realtime-0.5B的实时文本转语音（TTS）模型，参数量为0.5B。
• 该模型具备极低延迟特性，从输入文本到发出第一个声音平均仅需300毫秒。
• 它能够一次性生成长达90分钟的连续音频，并支持最多4个不同角色的对话模拟。
• 模型能根据文本语义自动识别并添加愤怒、激动、歉意等情绪，无需手动标注。
• 该模型已在GitHub上根据MIT协议开源，支持商用，并且因其轻量化设计，可在手机和边缘设备上运行。

相关链接：

• 微软开源实时语音模型VibeVoice-Realtime-0.5B，300ms实时开声，90分钟长音频都不喘！

Microsoft Launches Lightweight Real-time Voice Model, Supports 90-Minute Ultra-Long Audio Generation

Brief:

• Microsoft has open-sourced VibeVoice-Realtime-0.5B, a real-time text-to-speech (TTS) model with 0.5 billion parameters. 🚀
• The model boasts extremely low latency, averaging just 300 milliseconds from text input to the first sound output.
• It can generate up to 90 minutes of continuous audio in a single go and supports dialogue simulation for up to 4 different characters.
• The model can automatically recognize and add emotions such as anger, excitement, or apology based on text semantics, without the need for manual labeling. 🎤
• The model is open-sourced on GitHub under the MIT license, supports commercial use, and its lightweight design allows it to run on mobile phones and edge devices. 💡

Related Links:

• Microsoft Open-Sources Real-time Voice Model VibeVoice-Realtime-0.5B: 300ms Real-time Activation, Seamless 90-Minute Long Audio Generation!