单词学习：![[2026041731881852.mp3]]

新闻播报：![[2026041761147964.mp3]]

谷歌开源 MedGemma 1.5 医疗 AI 模型，突破 2D 局限全面支持 3D 影像解析

简报：

• 谷歌宣布开源医疗 AI 模型 MedGemma 1.5，该版本的核心突破在于打破传统 2D 图像局限，原生支持 CT 和 MRI 等三维医疗扫描数据的分析。
• 模型功能实现全面升级，除支持处理数十个切片的立体影像外，还具备病理学全视野数字切片分析、胸部 X 光病变精准标定、多时间点病情对比，以及大幅提升的 PDF 电子病历解析能力。
• 性能方面，在保持 40 亿参数量不变的前提下，相较于上一代产品，其 3D MRI 病情分类准确率提升 11%，胸部 X 光解剖定位交并比增加 35%，电子病历问答准确率提升 22%。
• 技术路线上，研发团队注入了涵盖放射学、皮肤科等领域的医学图文数据，设计了针对 3D CT 扫描的序列切分预处理方法，并引入领域专家模型进行知识蒸馏。
• 谷歌明确该模型并非开箱即用的临床决策工具，而是定位为供开发者微调的基础资源；同时在演进为“医疗通才”的过程中，模型在部分小众旧视觉问答基准上出现极轻微退步。

相关链接：

• 谷歌开源医疗 AI 模型 MedGemma 1.5，支持 3D 影像读取
• MedGemma 1.5 论文地址

Google Open-Sources MedGemma 1.5 Medical AI Model, Breaking Through 2D Limitations with Full Support for 3D Imaging Analysis 🩺📊

Brief:

• Google has announced the open-source release of its medical AI model MedGemma 1.5. The core breakthrough of this version lies in overcoming the limitations of traditional 2D imaging, with native support for analyzing 3D medical scan data such as CT and MRI.
• The model’s capabilities have been comprehensively upgraded. In addition to processing volumetric images composed of dozens of slices, it now supports whole-slide digital pathology analysis, precise lesion localization in chest X-rays, multi-timepoint disease comparison, and significantly improved PDF electronic medical record parsing.
• In terms of performance, while maintaining the same 4 billion parameter scale, the new version delivers notable gains over the previous generation: 3D MRI disease classification accuracy has improved by 11%, chest X-ray anatomical localization IoU has increased by 35%, and electronic medical record question-answering accuracy has risen by 22%.
• On the technical side, the research team incorporated medical image-text data spanning radiology, dermatology, and other specialties, designed a sequence-splitting preprocessing method tailored for 3D CT scans, and introduced domain expert models for knowledge distillation.
• Google has made it clear that the model is not intended as an out-of-the-box clinical decision-making tool, but rather as a foundational resource for developers to fine-tune. At the same time, as it evolves toward becoming a “medical generalist,” the model shows very slight regressions on some niche legacy visual question-answering benchmarks.

Related links:

• Google Open-Sources Medical AI Model MedGemma 1.5 with Support for 3D Imaging
• MedGemma 1.5 Paper

OpenAI Codex 大幅升级：新增自主操控 Mac 桌面与长线任务执行能力

简报：

• OpenAI 对 Codex AI 编程助手进行重磅更新，核心方向涵盖 Mac 桌面应用操控、多智能体并行运行以及长期任务执行。
• Codex 新增直接操作 Mac 桌面应用的能力，拥有独立光标，可全程在无人工介入下自主查看屏幕、点击和输入文字，且多个智能体可并行运行互不打扰。
• 在任务连续性上，Codex 增加记忆功能以保存用户偏好与工作流，支持无缝恢复暂停的任务，并能自主调度未来数天至数周的工作计划。
• 浏览器体验得到强化，应用内浏览器支持页面批注，OpenAI 还透露未来将推出完整的浏览器操控能力。
• 本次更新还集成了 gpt-image-1.5 模型用于生成概念图和原型，并新增超 90 款插件，大幅扩充了应用集成与 MCP 服务器接入生态。

相关链接：

• OpenAI 更新 Codex：支持 Mac 桌面操控

OpenAI Codex Gets Major Upgrade: Adds Autonomous Mac Desktop Control and Long-Running Task Execution

Brief:

• OpenAI has released a major update to its Codex AI coding assistant, with core improvements covering Mac desktop app control, parallel multi-agent operation, and long-running task execution.
• Codex now gains the ability to directly operate Mac desktop applications. It has its own independent cursor and can autonomously view the screen, click, and enter text without human intervention, while multiple agents can run in parallel without interfering with one another.
• For task continuity, Codex has added memory features to save user preferences and workflows, support seamless resumption of paused tasks, and autonomously schedule work plans over the coming days or weeks.
• The browser experience has been enhanced. The in-app browser now supports page annotations, and OpenAI also said it plans to introduce full browser control capabilities in the future.
• This update also integrates the gpt-image-1.5 model for generating concept images and prototypes, and adds more than 90 new plugins, greatly expanding the ecosystem for app integrations and MCP server access. 🚀

Related link:

• OpenAI updates Codex with support for Mac desktop control

OpenAI Codex 迎重磅更新：新增Mac桌面操控、多智能体并行与长期任务记忆

简报：

• OpenAI于今日（4月17日）对Codex AI编程助手进行核心升级，主打Mac桌面应用操控、多智能体并行运行以及更长时间维度的任务执行能力。
• Codex现已获得直接操作Mac桌面应用的能力，拥有独立光标，可自主查看、点击和输入；支持多个智能体同时并行运行，且互不干扰用户的正常工作流。
• 新增任务连续性与记忆功能，可保存用户偏好和技术栈信息；任务暂停后能无缝恢复，并可自主调度未来数周的工作计划，主动提出推进建议。
• 浏览器体验获得强化，应用内浏览器支持用户在页面直接添加批注；OpenAI表示未来将上线完整的浏览器自主操控与截图检验功能。
• 其他更新包括：集成gpt-image-1.5模型生成产品概念图和界面原型，侧边栏支持预览PDF和电子表格等文件，同时插件生态新增逾90款插件。

相关链接：

• OpenAI更新Codex：支持Mac桌面操控、多智能体并行与长期任务执行

OpenAI Codex Gets Major Update: Adds Mac Desktop Control, Multi-Agent Parallel Execution, and Long-Term Task Memory

Brief:

• OpenAI today, April 17, rolled out a core upgrade to Codex, its AI programming assistant, focusing on Mac desktop app control, multi-agent parallel execution, and longer-horizon task execution capabilities.
• Codex can now directly operate Mac desktop apps. It has its own independent cursor and can autonomously view, click, and type. It also supports multiple agents running in parallel at the same time, without interfering with the user’s normal workflow.
• New task continuity and memory features can save user preferences and technical stack information. After a task is paused, Codex can resume seamlessly, independently schedule work plans for the coming weeks, and proactively suggest next steps.
• The browser experience has been enhanced. The in-app browser now supports adding annotations directly on webpages. OpenAI says full autonomous browser control and screenshot verification features will be launched in the future. 🌐
• Other updates include integration with the gpt-image-1.5 model to generate product concept images and UI prototypes, sidebar support for previewing files such as PDFs and spreadsheets, and an expanded plugin ecosystem with more than 90 new plugins. 🚀

Related link:

• OpenAI Updates Codex With Support for Mac Desktop Control, Multi-Agent Parallel Execution, and Long-Term Task Execution

Anthropic发布Claude Opus 4.7：工程与视觉能力升级且定价不变

简报：

• Anthropic正式推出最新人工智能模型Claude Opus 4.7，相较于4.6版本，在高级软件工程、复杂任务处理和指令执行方面均有显著提升，并具备自我验证能力。
• 视觉处理表现增强，支持识别高达2576像素长边的图像，提升了在创建高质量幻灯片和文档等专业任务中的创造力与质量。
• 安全与记忆功能升级，内置网络安全防护以自动拦截高风险请求，同时改进内存利用率，能够有效记住长时间会话中的重要笔记。
• 价格与前代保持一致（输入每百万token 5美元，输出每百万token 25美元），但因指令执行力增强，用户需适度调整提示信息以防意外结果。
• 该模型现已全面上线涵盖Claude产品及API、Amazon Bedrock、Google Cloud Vertex AI以及Microsoft Foundry等多个平台，并推出了细致的难度控制和任务预算功能。

相关链接：

• Anthropic 发布全新 Claude Opus 4.7 模型

Anthropic Releases Claude Opus 4.7: Engineering and Vision Capabilities Upgraded, Pricing Unchanged

Brief:

• Anthropic has officially launched its latest AI model, Claude Opus 4.7. Compared with version 4.6, it delivers notable improvements in advanced software engineering, complex task handling, and instruction following, while also adding self-verification capabilities.
• Vision processing has been enhanced, with support for recognizing images with a long edge of up to 2,576 pixels. This improves creativity and quality in professional tasks such as creating high-quality slides and documents. 🖼️
• Safety and memory features have been upgraded. Built-in cybersecurity protections can automatically block high-risk requests, while improved memory usage helps the model effectively retain important notes across long-running conversations.
• Pricing remains unchanged from the previous generation, at $5 per million input tokens and$25 per million output tokens. However, due to stronger instruction-following capabilities, users should adjust prompts appropriately to avoid unintended results.
• The model is now fully available across multiple platforms, including Claude products and APIs, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. It also introduces fine-grained difficulty controls and task budget features. 🚀

Related link:

• Anthropic Releases New Claude Opus 4.7 Model

NVIDIA发布首个开源量子AI模型“Ising”，大幅提升校准与纠错效率

简报：

• NVIDIA宣布推出全球首个开源量子AI模型“NVIDIA Ising”，专门用于解决量子处理器校准和量子纠错两大关键难题。
• 模型包含拥有350亿参数的视觉语言模型Ising Calibration，可将测量数据的处理时间从数天大幅缩短至仅数小时。
• 模型还包含Ising Decoding，在实时解码过程中提供了比行业基准pyMatching高出2.5倍的速度和3倍的准确率，且所需训练数据仅为十分之一。
• “NVIDIA Ising”目前已被IonQ、哈佛大学和费米国家加速器实验室等多家机构应用，为构建混合量子-经典系统提供支持。
• 该模型的命名源于统计力学中的“Ising 模型”，其发布进一步丰富了NVIDIA涵盖智能体、物理AI、辅助驾驶及生物医学等领域的开放模型体系。

相关链接：

• NVIDIA 推出全球首个开源量子 AI 模型

NVIDIA Releases First Open-Source Quantum AI Model “Ising,” Significantly Improving Calibration and Error-Correction Efficiency

Briefing:

• NVIDIA has announced the launch of “NVIDIA Ising,” the world’s first open-source quantum AI model, designed to address two major challenges in quantum processors: calibration and quantum error correction.
• The model includes Ising Calibration, a vision-language model with 35 billion parameters that can reduce the processing time for measurement data from several days to just a few hours.
• The model also includes Ising Decoding, which delivers 2.5 times the speed and 3 times the accuracy of the industry benchmark pyMatching during real-time decoding, while requiring only one-tenth of the training data.
• “NVIDIA Ising” is already being used by several institutions, including IonQ, Harvard University, and Fermilab, supporting the development of hybrid quantum-classical systems.
• The model is named after the “Ising model” in statistical mechanics. Its release further expands NVIDIA’s open model ecosystem across areas such as agents, physical AI, assisted driving, and biomedicine.

Related link:

• NVIDIA Launches the World’s First Open-Source Quantum AI Model

我国实现两大突破，牵头立项全球首个具身智能国际标准

简报：

• 4月17日报道，我国在国际标准化组织成功立项具身智能领域全球首项国际标准《人形机器人数据集》。
• 我国在此次立项中实现了两个历史性的“零的突破”：一是人形机器人国际标准制定，二是机器人国际标准工作组建设，并推动成立首个由我国专家担任召集人的工作组。
• 此次立项的国际标准参考了我国在研的人形机器人数据集系列国家标准内容，全球多国专家将携手参与后续标准研制。
• 数据显示，截至2025年底我国已成为全球第一大人形机器人市场，出货量达1.44万台，占据全球总量的84.7%。
• 作为顶层设计铺垫，今年2月我国已正式发布《人形机器人与具身智能标准体系（2026版）》，涵盖基础共性、类脑与智算、肢体与部组件、整机与系统、应用、安全伦理6个部分。

相关链接：

• 具身智能领域首个国际标准由我国牵头成功立项
• 人形机器人标准化技术委员会委员名单

China Achieves Two Major Breakthroughs, Leads the Launch of the World’s First International Standard for Embodied Intelligence

Briefing:

• Reported on April 17, China has successfully initiated the world’s first international standard in the field of embodied intelligence under the International Organization for Standardization, titled Humanoid Robot Datasets.
• China achieved two historic firsts in this project: the first international standard for humanoid robots, and the first international standards working group for robotics, with China also推动 the establishment of the first working group convened by a Chinese expert.
• The newly initiated international standard draws on China’s national standards under development for humanoid robot dataset series, with experts from multiple countries set to participate in the subsequent drafting work.
• Data shows that by the end of 2025, China had become the world’s largest humanoid robot market, with shipments reaching 14,400 units and accounting for 84.7% of the global total.
• As part of the top-level design groundwork, China officially released the Standard System for Humanoid Robots and Embodied Intelligence (2026 Edition) in February this year, covering six areas: basic common technologies, brain-inspired intelligence and intelligent computing, limbs and components, complete machines and systems, applications, and safety ethics. 🤖

Related Links:

• China Successfully Leads the Launch of the First International Standard in the Field of Embodied Intelligence
• List of Members of the Humanoid Robot Standardization Technical Committee