单词学习：![[2026041775446978.mp3]]

新闻播报：![[2026041768328015.mp3]]

OpenAI拟超200亿美元采购Cerebras服务器以减少对英伟达依赖

简报：

消息称OpenAI同意在未来三年内向Cerebras支付超过200亿美元。
这笔巨额资金将用于使用由Cerebras芯片驱动的服务器集群器。
OpenAI此次合作的主要目的是通过多元化硬件布局来显著降低计算成本，同时减少对英伟达芯片的依赖。

相关链接：

豪掷200亿美元，OpenAI据称将与Cerebras达成重磅协议

OpenAI Reportedly Plans to Purchase More Than $20 Billion in Cerebras Servers to Reduce Reliance on Nvidia 🚀

Brief:

Reports say OpenAI has agreed to pay more than $20 billion to Cerebras over the next three years.
This massive funding will be used to access server clusters powered by Cerebras chips.
The primary goal of this partnership is to significantly lower computing costs through a more diversified hardware strategy, while also reducing OpenAI’s dependence on Nvidia chips 🤖💡

谷歌开源40亿参数医疗AI MedGemma 1.5，全面升级3D多维度影像分析能力

简报：

谷歌开源医疗AI基础模型MedGemma 1.5，核心突破在于克服传统2D图像局限，原生支持CT和MRI三维扫描数据以及多时间点病情对比分析。
此外，模型还支持病理学全视野数字切片、胸部X光病变边界框精准标注，大幅提升对PDF格式电子病历和化验单的结构化数据提取理解能力。
在保持40亿参数计算效率的前提下，新版本在3D MRI分类准确率上提升11%，病理图像宏观F1分数提升47%，电子病历问答准确率提升22%。
训练过程中使用了放射学、皮肤科等大量医学图文数据，设计了最高85张切分序列图像的预处理方法，并引入专家模型进行知识蒸馏。
谷歌将其明确界定为供开发者微调的底层资源，目前并非开箱即用的临床决策工具，实际落地仍需具体场景的专项训练。

相关链接：

Google Open-Sources 4-Billion-Parameter Medical AI MedGemma 1.5, Fully Upgrading Its 3D Multi-Dimensional Imaging Analysis Capabilities 🏥🧠

Brief:

Google has open-sourced the medical AI foundation model MedGemma 1.5. Its core breakthrough lies in overcoming the limitations of traditional 2D images, with native support for 3D CT and MRI scan data as well as comparative analysis of disease progression across multiple time points.
In addition, the model supports whole-slide digital pathology images and precise bounding-box annotation of lesions in chest X-rays, while significantly improving its ability to extract and understand structured data from PDF-based electronic medical records and lab reports.
While maintaining the computational efficiency of a 4-billion-parameter model, the new version improves 3D MRI classification accuracy by 11%, boosts the macro F1 score for pathology images by 47%, and increases electronic medical record question-answering accuracy by 22%.
During training, it used large volumes of medical image-text data from radiology, dermatology, and other fields, introduced a preprocessing method capable of handling up to 85 segmented sequential images, and incorporated expert models for knowledge distillation.
Google has clearly defined it as a foundational resource for developers to fine-tune. It is not currently an out-of-the-box clinical decision-making tool, and real-world deployment will still require task-specific training for particular use cases. 📄

Google AI Studio接入AI订阅服务，按请求付费解锁所有模型与智能体权限

简报：

Google 在 AI Studio 上逐步推出对 Google AI 计划的支持，用户可直接关联订阅，获得更高的使用限额及更流畅的文本、图像、音频等多模态模型访问体验。
新增的“按请求付费（Pay-per-request）”选项目前提供对“所有模型和代理（智能体）”的完整访问权限，而 Google AI 订阅计划本身暂不包含此项完整功能。
按请求付费模式让用户能够无限制探索最新模型，而订阅计划则侧重于稳定的限额提升和跨产品整合。
更新内容暗示 Google 可能正在测试或逐步开放基于 Gemini 等模型的智能体（Agent）能力，以帮助开发者通过自然语言快速构建应用。
目前有用户反馈，在切换订阅时偶尔会出现确认流程卡顿的问题。

相关链接：

Google AI Studio 直接支持 AI 订阅计划，按请求付费独享“所有模型+智能体”？

Google AI Studio Integrates AI Subscription Services, Unlocking Access to All Models and Agents with Pay-per-Request 🚀

Brief:

Google is gradually rolling out support for the Google AI plan in AI Studio, allowing users to link their subscriptions directly for higher usage limits and a smoother multimodal access experience across text, image, audio, and more.
The newly added “Pay-per-request” option currently provides full access to “all models and agents,” while the Google AI subscription plan itself does not yet include this complete functionality.
The pay-per-request model enables users to explore the latest models without restriction, while the subscription plan focuses more on stable quota increases and cross-product integration.
This update suggests that Google may be testing or gradually opening up Agent capabilities based on models such as Gemini, helping developers build applications more quickly through natural language. 🤖
Some users have reported occasional confirmation flow lag when switching subscriptions.

Anthropic推出Claude Opus 4.7性能大幅提升，同步实施强制实名验证引争议

简报：

Anthropic今日正式发布Claude Opus 4.7模型，在高级软件工程、复杂任务处理及指令执行方面较上一代有显著提升，并具备结果自我验证能力。
新模型视觉处理能力增强，支持识别最高2,576像素长边的图像，同时内置网络安全防护措施，可自动检测并拦截高风险请求。
Opus 4.7优化了长会话的内存利用率，推出更细致的难度控制和任务预算功能，定价维持不变（输入 $5/百万token，输出$ 25/百万token）。
目前Opus 4.7已在Claude产品及API、Amazon Bedrock、Google Cloud Vertex AI和Microsoft Foundry等平台全面上线。
此外，Claude近期实施强制实名验证（KYC），要求用户手持实体政府证件实时拍照，不支持复印或数字件，引发用户对数据隐私及封号风险的担忧。
Anthropic严格设定最低18岁的使用年龄限制，已有未成年开发者因触发限制被系统封号并退款。

相关链接：

Anthropic Launches Claude Opus 4.7 with Major Performance Gains, While Mandatory Real-Name Verification Sparks Controversy

Briefing:

Anthropic officially launched the Claude Opus 4.7 model today, delivering notable improvements over the previous generation in advanced software engineering, complex task handling, and instruction following, while also gaining the ability to self-verify results. 🚀
The new model features stronger visual processing capabilities, supporting image recognition with a maximum long side of 2,576 pixels, and includes built-in cybersecurity safeguards that can automatically detect and block high-risk requests.
Opus 4.7 also improves memory efficiency in long conversations, introduces more granular difficulty control and task budget features, and keeps pricing unchanged at $5 per million input tokens and$ 25 per million output tokens.
Opus 4.7 is now fully available across Claude products and API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
In addition, Claude has recently rolled out mandatory real-name verification, requiring users to take a live photo while holding a physical government-issued ID. Copies and digital versions are not accepted, raising concerns over data privacy and the risk of account bans. ⚠️
Anthropic also enforces a strict minimum age requirement of 18, and some underage developers have reportedly had their accounts banned and refunded after triggering the restriction.

阿里正式开源低耗MoE模型Qwen3.6-35B-A3B，以30亿激活参数突破智能体编程能力

简报：

阿里巴巴千问大模型团队于2026年4月16日正式开源稀疏混合专家（MoE）模型 Qwen3.6-35B-A3B，标志着轻量级模型在智能体编程领域实现关键突破。
该模型总参数量为350亿，运行时激活参数仅为30亿，以极低的计算成本在核心编程基准测试中超越了270亿参数的稠密模型 Qwen3.5-27B。
模型逻辑推理与智能体协同能力可比肩 Gemma4-31B 等大体量模型，并在空间智能与视觉感知方面表现优异，部分多模态指标达到 Claude Sonnet4.5 水平。
Qwen3.6-35B-A3B 目前已集成至 Qwen Studio，并在阿里云百炼平台以 qwen3.6-flash 之名提供 API 服务，可无缝适配 OpenClaw 等主流 AI 编程助手。
此外，通义千问关联公司通义云启（杭州）信息技术有限公司近日注册资本由1亿元增至2亿元翻倍，并新增杭州通义实验室为股东，与阿里云共同持股。

相关链接：

Alibaba Officially Open-Sources the Low-Cost MoE Model Qwen3.6-35B-A3B, Achieving a Breakthrough in Agentic Coding with 3 Billion Active Parameters 🚀

Briefing:

On April 16, 2026, Alibaba’s Qwen large model team officially open-sourced the sparse Mixture-of-Experts (MoE) model Qwen3.6-35B-A3B, marking a major breakthrough for lightweight models in the field of agentic coding.
The model has a total of 35 billion parameters, but activates only 3 billion parameters at runtime. With extremely low computational cost, it surpasses the 27-billion-parameter dense model Qwen3.5-27B on core coding benchmark tests.
Its logical reasoning and agent collaboration capabilities are comparable to larger-scale models such as Gemma4-31B, while also delivering strong performance in spatial intelligence and visual perception. On some multimodal benchmarks, it reaches the level of Claude Sonnet4.5 🤖
Qwen3.6-35B-A3B has now been integrated into Qwen Studio and is also available through API services on Alibaba Cloud’s Bailian platform under the name qwen3.6-flash, enabling seamless compatibility with mainstream AI coding assistants such as OpenClaw.
In addition, Tongyi Yunqi (Hangzhou) Information Technology Co., Ltd., an affiliated company of Tongyi Qianwen, recently doubled its registered capital from RMB 100 million to RMB 200 million, and added Hangzhou Tongyi Lab as a new shareholder, jointly held with Alibaba Cloud 📈

字节跳动面向全球开启前沿技术博士生校招，并加码AI人才激励

简报：

字节跳动正式启动前沿技术领域人才校园招聘项目，面向全球优秀技术人才开放全职及实习岗位。
招聘对象主要为博士群体，全职岗位面向2026年9月至2027年8月毕业的博士生，实习岗位面向2026年9月及以后毕业的在校博士生。
顶会顶刊论文作者、专利拥有者、国际竞赛获奖者及有重大影响力技术落地成果的人才将获优先考量。
开放岗位方向涵盖大模型应用、搜索/推荐/广告、安全/AI Safety、AIGC、AI Coding、视频架构及AI for Science等多个前沿技术领域。
工作地点覆盖国内五城（北京、上海、深圳、杭州、成都）及海外五城（新加坡、悉尼、圣何塞、西雅图、圣地亚哥）。
为增强对顶尖AI人才的吸引力与激励，字节跳动同步启动了针对大模型业务核心人员的首轮“豆包股”回购，回购价较授予价上涨约30%。

相关链接：

ByteDance Launches Global Campus Recruitment for PhD Talent in Frontier Technologies and Steps Up AI Incentives 🚀

Brief:

ByteDance has officially launched a campus recruitment program for talent in frontier technology fields, opening both full-time and internship positions to outstanding technical talent worldwide.
The program mainly targets PhD candidates. Full-time roles are open to PhD students graduating between September 2026 and August 2027, while internship roles are open to current PhD students graduating in September 2026 or later.
Priority will be given to candidates who have published papers at top conferences or in leading journals, hold patents, have won international competitions, or have delivered highly impactful technical achievements.
Open roles cover a wide range of frontier technology areas, including large model applications, search/recommendation/advertising, security/AI Safety, AIGC, AI Coding, video architecture, and AI for Science.
Job locations span five cities in China (Beijing, Shanghai, Shenzhen, Hangzhou, and Chengdu) and five overseas cities (Singapore, Sydney, San Jose, Seattle, and San Diego).
To further strengthen its appeal to top AI talent and improve incentives, ByteDance has also launched the first round of repurchases of Doubao shares for core personnel in its large model business, with the repurchase price about 30% higher than the grant price 📈🤖

目录

OpenAI拟超200亿美元采购Cerebras服务器以减少对英伟达依赖

OpenAI Reportedly Plans to Purchase More Than $20 Billion in Cerebras Servers to Reduce Reliance on Nvidia 🚀

谷歌开源40亿参数医疗AI MedGemma 1.5，全面升级3D多维度影像分析能力

Google Open-Sources 4-Billion-Parameter Medical AI MedGemma 1.5, Fully Upgrading Its 3D Multi-Dimensional Imaging Analysis Capabilities 🏥🧠

Google AI Studio接入AI订阅服务，按请求付费解锁所有模型与智能体权限

Google AI Studio Integrates AI Subscription Services, Unlocking Access to All Models and Agents with Pay-per-Request 🚀

Anthropic推出Claude Opus 4.7性能大幅提升，同步实施强制实名验证引争议

Anthropic Launches Claude Opus 4.7 with Major Performance Gains, While Mandatory Real-Name Verification Sparks Controversy

阿里正式开源低耗MoE模型Qwen3.6-35B-A3B，以30亿激活参数突破智能体编程能力

Alibaba Officially Open-Sources the Low-Cost MoE Model Qwen3.6-35B-A3B, Achieving a Breakthrough in Agentic Coding with 3 Billion Active Parameters 🚀

字节跳动面向全球开启前沿技术博士生校招，并加码AI人才激励

ByteDance Launches Global Campus Recruitment for PhD Talent in Frontier Technologies and Steps Up AI Incentives 🚀