Runway推出世界模型GWM-1，进军通用人工智能赛道

简报：

• AI视频公司Runway发布其首个通用世界模型（GWM-1），旨在通过预测视频像素来构建一个可理解物理规律的动态仿真环境，与谷歌、OpenAI等巨头竞争下一代人工智能基础设施。
• GWM-1初期将以三个分支落地：用于生成交互式动态世界的GWM-Worlds、辅助机器人训练的GWM-Robotics，以及生成数字人的GWM-Avatars，最终目标是融合成统一模型。
• Runway同时升级了其Gen4.5视频生成模型，新增了原生音频生成、一分钟多镜头视频合成、保持角色一致性等功能，并已向所有付费用户开放。

相关链接：

• Runway发布通用世界模型GWM-1，用像素预测构建可推理的虚拟世界

Runway Launches World Model GWM-1, Entering the General AI Race 🚀

Brief:

• AI video company Runway has released its first General World Model (GWM-1), aiming to build a dynamic simulation environment that understands physical laws by predicting video pixels, competing with giants like Google and OpenAI for the next generation of AI infrastructure.
• GWM-1 will initially launch with three branches: GWM-Worlds for generating interactive dynamic worlds, GWM-Robotics for assisting robot training, and GWM-Avatars for generating digital humans. The ultimate goal is to merge these into a unified model.
• Runway has also upgraded its Gen4.5 video generation model, adding new features such as native audio generation, one-minute multi-shot video synthesis, and character consistency preservation. This upgrade is now available to all paying users. ✨

Related Links:

• Runway releases General World Model GWM-1, building an inferable virtual world by predicting pixels 🔗

蚂蚁集团发布并开源LLaDA2.0，业界首个千亿参数扩散语言模型

简报：

• 蚂蚁技术研究院于12月12日宣布推出LLaDA2.0系列离散扩散大语言模型（dLLM），宣称是“业内首个100B扩散语言模型”。
• LLaDA2.0系列包含16B和100B两个版本，首次将Diffusion模型的参数规模扩展到千亿量级。
• 官方称，该模型在代码、数学和智能体任务上表现超越同级自回归（AR）模型，并实现了2.1倍的推理加速。
• LLaDA2.0的16B和100B版本模型权重及相关训练代码均已在Huggingface上开源。

相关链接：

• 蚂蚁开源业内首个 100B 扩散语言模型 LLaDA2.0
• LLaDA 2.0 开源项目地址

Ant Group Releases and Open-Sources LLaDA2.0, the Industry's First Hundred-Billion-Parameter Diffusion Language Model 🚀

Briefing:

• Ant Tech Research announced on December 12 the launch of the LLaDA2.0 series of discrete diffusion large language models (dLLM), claiming it to be "the industry's first 100B diffusion language model."
• The LLaDA2.0 series includes 16B and 100B versions, marking the first time Diffusion models have scaled to the hundred-billion-parameter level.
• According to official statements, the model outperforms comparable auto-regressive (AR) models in code, mathematics, and agent tasks, and achieves a 2.1x inference acceleration.
• The model weights for both the 16B and 100B versions of LLaDA2.0, along with relevant training code, have been open-sourced on Huggingface. 💡🌐

Related Links:

• Ant Open-Sources LLaDA2.0, the Industry's First 100B Diffusion Language Model
• LLaDA 2.0 Open Source Project Address

谷歌发布Gemini研究代理并开放API，与OpenAI展开正面竞争

简报：

• 为应对OpenAI发布的GPT-5.2，谷歌推出其升级版AI研究代理Gemini Deep Research，基于最新的Gemini 3 Pro模型构建。
• 该代理专为处理海量信息与超长上下文的复杂研究任务设计，旨在降低AI“幻觉”，提升事实准确性。
• 谷歌首次通过全新的Interactions API向开发者开放该能力，允许将高级研究功能嵌入第三方应用，并支持处理上传文件与结构化输出。
• 为评测AI代理的研究能力，谷歌同步推出了新的开源基准测试DeepSearchQA，用于衡量多步骤复杂检索任务的表现。
• 该能力未来将被集成到Google Search、NotebookLM、Google Finance及Gemini App等谷歌产品中。

相关链接：

• 对抗 OpenAI GPT-5.2：谷歌祭出 Gemini Deep Research 深度研究智能体并首次向开发者开放，DeepSearchQA 因果链测试基准重磅开源
• 谷歌紧急上线Gemini Deep Research迎战GPT-5.2，AI“研究代理”大战一触即发

Google Launches Gemini Research Agent and Opens API, Entering Direct Competition with OpenAI 🚀

Brief:

• In response to OpenAI's GPT-5.2 release, Google unveils its upgraded AI research agent, Gemini Deep Research 🧠, built on the latest Gemini 3 Pro model.
• This agent is specifically designed to handle complex research tasks involving vast amounts of information and ultra-long contexts, aiming to reduce AI "hallucinations" and enhance factual accuracy.
• Google is making this capability available to developers for the first time via the new Interactions API 🛠️, allowing advanced research features to be embedded into third-party applications and supporting the processing of uploaded files and structured output.
• To evaluate the research capabilities of AI agents, Google has simultaneously launched a new open-source benchmark, DeepSearchQA, designed to measure performance in multi-step complex retrieval tasks.
• This capability will be integrated into Google products such as Google Search, NotebookLM, Google Finance, and the Gemini App.

Related Links:

• 对抗 OpenAI GPT-5.2：谷歌祭出 Gemini Deep Research 深度研究智能体并首次向开发者开放，DeepSearchQA 因果链测试基准重磅开源
• 谷歌紧急上线Gemini Deep Research迎战GPT-5.2，AI“研究代理”大战一触即发

谷歌发布Gemini深度研究代理，开放API迎战OpenAI

简报：为应对OpenAI的GPT-5.2，谷歌发布了基于Gemini 3 Pro的深度研究智能体Gemini Deep Research，并通过全新的Interactions API首次向开发者开放其高级研究能力，同时开源了用于衡量复杂检索任务的评测基准DeepSearchQA。

相关链接：

• 对抗 OpenAI GPT-5.2：谷歌祭出 Gemini Deep Research 深度研究智能体并首次向开发者开放，DeepSearchQA 因果链测试基准重磅开源

Google Releases Gemini Deep Research Agent, Opens API to Challenge OpenAI

Brief: In a strategic move to counter OpenAI's GPT-5.2, Google has unveiled Gemini Deep Research, a powerful deep research agent built on Gemini 3 Pro. 🧠 For the very first time, its advanced research capabilities are being opened to developers through the innovative Interactions API. Concurrently, Google has open-sourced DeepSearchQA, a new benchmark designed to evaluate complex retrieval tasks. This move marks a significant step in the AI race! 🚀✨

Related Link:

• Challenging OpenAI GPT-5.2: Google Launches Gemini Deep Research Agent, Opens to Developers for the First Time, and Open-Sources DeepSearchQA Causal Chain Benchmark

迪士尼与OpenAI达成合作，Sora将可生成漫威与星球大战等角色视频

简报：

• 迪士尼与OpenAI宣布达成一项为期三年的合作协议，迪士尼将向OpenAI进行10亿美元的股权投资。
• OpenAI获准使用超过200个来自迪士尼、皮克斯、漫威和星球大战的角色，用于Sora视频生成和ChatGPT图像生成。
• 合作生成的授权短视频和艺术作品将在Disney+平台上展出。
• 该协议不包括真人角色的肖像和声音，意味着漫威、星球大战等真人角色将仅以动画或插画形式出现。
• 此次合作是OpenAI在Sora2视频工具发布后，为应对版权争议而与内容所有者签订许可协议的举措之一。

相关链接：

• 迪士尼与 OpenAI 达成协议，授权使用热门角色创作视频与图像

Disney and OpenAI Partner, Sora to Generate Videos of Marvel, Star Wars, and Other Characters 🎬

Brief:

• Disney and OpenAI announced a three-year collaboration agreement, with Disney making a dJyH6kDkfC5RlM16.answerText billion equity investment in OpenAI. 🤝
• OpenAI is authorized to use over 200 characters from Disney, Pixar, Marvel, and Star Wars for Sora video generation and ChatGPT image generation. 🪄
• Licensed short videos and artworks generated through the collaboration will be showcased on the Disney+ platform.
• The agreement does not include the likeness and voices of live-action characters, meaning live-action Marvel, Star Wars, and other characters will only appear in animated or illustrated forms.
• This partnership is one of OpenAI's steps to sign licensing agreements with content owners to address copyright concerns following the release of its Sora2 video tool.

Related Link:

• Disney Reaches Agreement with OpenAI, Authorizing Use of Popular Characters for Video and Image Creation

因应AI需求，台积电熊本二厂拟升级4nm制程并暂停施工

简报：

• 据日经亚洲报道，为应对激增的人工智能（AI）芯片需求，台积电正考虑将其日本熊本二厂的生产工艺从原定的6纳米与7纳米，升级为更先进的4纳米。
• 此举可能导致工厂重新设计并延后原定于2027年的量产时程。报道称，熊本二厂的施工已暂时停止，多数重型设备已撤离工地。
• 同时，台积电也已暂缓为熊本一厂增添设备的计划，并告知多家供应商2026年全年都无须新增设备。
• 台积电方面回应称，日本项目正在推进中，公司正与合作伙伴商讨详细的施工执行计划。

相关链接：

• 台积电熊本二厂传考虑产「更先进」4奈米晶片！　因应AI需求

TSMC Kumamoto Fab 2 Considers 4nm Process Upgrade, Halts Construction Amid AI Demand 🚀

Briefing:

• According to Nikkei Asia, TSMC is reportedly considering upgrading the production process at its second factory in Kumamoto, Japan, from the originally planned 6nm and 7nm to a more advanced 4nm to meet surging demand for artificial intelligence (AI) chips.
• This move could lead to a factory redesign and delay the previously scheduled mass production timeline of 2027. Reports indicate that construction at Kumamoto Fab 2 has been temporarily halted, with most heavy equipment removed from the site.
• Concurrently, TSMC has also reportedly paused plans to add equipment to Kumamoto Fab 1 and informed multiple suppliers that no new equipment will be needed throughout 2026.
• TSMC responded by stating that its Japan projects are proceeding as planned, and the company is discussing detailed construction and execution plans with its partners. 🏗️💡

Related Link:

• TSMC Kumamoto Fab 2 Reportedly Considering "More Advanced" 4nm Chips! In Response to AI Demand

黑芝麻智能与元戎启行联手，共创高阶智驾解决方案

简报：

• 12月12日，芯片公司黑芝麻智能与自动驾驶公司元戎启行宣布达成深度合作，双方将建立资本与业务的双重纽带。
• 合作旨在结合黑芝麻智能在芯片技术和元戎启行在辅助驾驶算法上的优势，基于黑芝麻智能下一代车规级高性能计算芯片，联合打造高性价比的高阶辅助驾驶解决方案。
• 双方目标是共同推进技术规模化落地，并探索在Robotaxi等新兴领域的应用场景。

相关链接：

• 黑芝麻智能与元戎启行达成深度合作，共推高阶辅助驾驶技术量产落地

Black Sesame Technologies and DeepMotion Join Forces to Co-Create Advanced Intelligent Driving Solutions

Briefing:

• On December 12, chip company Black Sesame Technologies and autonomous driving company DeepMotion announced a deep strategic partnership, establishing dual ties in capital and business. 🤝
• The collaboration aims to combine Black Sesame Technologies' strengths in chip technology with DeepMotion's expertise in assisted driving algorithms. Based on Black Sesame Technologies' next-generation automotive-grade high-performance computing chips, they will jointly develop cost-effective advanced assisted driving solutions. 🚀
• Both parties aim to jointly promote the mass adoption of these technologies and explore application scenarios in emerging fields such as Robotaxi. 💡

Related Link:

• Black Sesame Technologies and DeepMotion Announce Deep Collaboration to Promote Mass Production of Advanced Assisted Driving Technology