GitHub Copilot全面升级为智能体，微软推出自主SWE智能体

简报：

• GitHub Copilot宣布全面接入智能体技术，推出Agent模式（预览版），具备自主迭代代码、识别错误并自动修复的能力
• 微软首次公开自主SWE智能体（项目代号Padawan），可像人类工程师一样主动解决编程问题
• 新功能Copilot Edits支持同时处理多个文件并提供代码修改建议，简化代码审查流程
• 新增视觉功能，支持上传截图、照片或图表生成相应界面和代码
• 微软CEO纳德拉表示此次升级将Copilot从"编程搭档"提升至"人类程序员水平"

相关链接：

• 微软官宣 All in 智能体、Copilot SWE Agent 首曝光
• GitHub Copilot全面智能化：微软重塑开发者编程体验
• 微软GitHub Copilot大升级：智能体全面接入，编程效率再飞跃！

GitHub Copilot Fully Upgraded to Intelligent Agent, Microsoft Launches Autonomous SWE Agent

Newsletter:

• GitHub Copilot announces full integration of intelligent agent technology, launching Agent Mode (preview version), capable of autonomously iterating code, identifying errors, and auto-fixing them 🛠️
• Microsoft unveils its first autonomous SWE Agent (project codename Padawan), which can proactively solve programming issues like a human engineer 💻
• New feature Copilot Edits supports simultaneous handling of multiple files and offers code modification suggestions, streamlining the code review process
• New visual feature allows users to upload screenshots, photos, or charts to generate corresponding interfaces and code 📸
• Microsoft CEO Nadella states that this upgrade elevates Copilot from a "programming partner" to the "level of a human programmer"

Related Links:

• Microsoft Officially Announces All-in on Agents, Copilot SWE Agent Debuts
• GitHub Copilot Fully Intelligent: Microsoft Reshapes Developer Programming Experience
• Microsoft GitHub Copilot Major Upgrade: Full Agent Integration, Programming Efficiency Soars Again!

DeepMind与陶哲轩合作推出AlphaEvolve，解决数学难题并优化芯片设计

简报：

• 谷歌DeepMind与数学家陶哲轩等合作推出通用科学智能体AlphaEvolve，该AI系统能自主生成和优化算法
• AlphaEvolve在数学领域取得突破，解决了300年历史的"接吻数问题"并改进了1969年的Strassen矩阵乘法算法
• 该系统已应用于谷歌数据中心和芯片设计，将Gemini训练时间缩短1%，FlashAttention提速32.5%
• AlphaEvolve采用生成-评估-进化循环架构，结合Gemini系列模型和自动化验证系统
• 陶哲轩表示该项目仍在进行中，未来将汇报更多进展，目前系统已开放早期测试申请

相关链接：

• 陶哲轩DeepMind梦幻联动，最强通用科学Agent来了！一口气解决芯片设计、矩阵乘法和300年几何难题
• 突破300年数学难题！陶哲轩出题，DeepMind通用科学AI智能体一夜屠龙
• 谷歌Agent深夜颠覆数学界！陶哲轩联手，解决300年难题或迎"AlphaGo时刻"

DeepMind Collaborates with Terence Tao to Launch AlphaEvolve, Tackling Mathematical Challenges and Optimizing Chip Design

Newsletter:

• Google DeepMind, in collaboration with mathematician Terence Tao and others, has unveiled AlphaEvolve, a universal scientific agent capable of autonomously generating and optimizing algorithms. 🧠
• AlphaEvolve has achieved breakthroughs in mathematics, solving the 300-year-old "kissing number problem" and improving the 1969 Strassen matrix multiplication algorithm.
• The system has been applied to Google data centers and chip design, reducing Gemini training time by 1% and accelerating FlashAttention by 32.5%. 🚀
• AlphaEvolve utilizes a generate-evaluate-evolve loop architecture, integrating the Gemini model series and an automated verification system.
• Terence Tao stated that the project is still ongoing, with more progress to be reported in the future. The system is currently open for early testing applications. 🌟

Related Links:

• Terence Tao and DeepMind's Epic Collaboration: The Ultimate Universal Scientific Agent Arrives! Solving Chip Design, Matrix Multiplication, and a 300-Year-Old Geometry Problem in One Go
• Breaking a 300-Year-Old Math Problem! Terence Tao Poses the Challenge, DeepMind's Universal Scientific AI Agent Slays the Dragon Overnight
• Google Agent Shakes Up the Math World Late at Night! Terence Tao Joins Forces, Solving a 300-Year-Old Problem May Usher in an "AlphaGo Moment"

AI教父Hinton承认高估AI取代放射科医生的能力

简报：

• Geoffrey Hinton承认2016年预测AI将在5年内取代放射科医生过于乐观，低估了医学复杂性和AI发展速度
• 梅奥诊所放射科医生数量从2016年260人增至400多人，增长55%，与Hinton预测相反
• AI目前主要作为辅助工具，梅奥诊所使用250多个AI模型提高诊断效率和准确性
• 放射科专家称AI是"第二双眼睛"，可处理重复性工作，但临床判断仍需医生

相关链接：

• AI 教父认错！Geoffrey Hinton改口：当年看走眼了，不该给放射科医生判"死缓"
• AI 教父认错！Geoffrey Hinton改口：当年看走眼了，不该给放射科医生判"死缓"

AI Godfather Hinton Admits Overestimating AI's Ability to Replace Radiologists

Newsletter:

• Geoffrey Hinton admits that his 2016 prediction of AI replacing radiologists within 5 years was overly optimistic, underestimating the complexity of medicine and the pace of AI development. 😮
• The number of radiologists at Mayo Clinic has increased from 260 in 2016 to over 400, a 55% growth, contrary to Hinton's forecast.
• AI currently serves mainly as a supportive tool, with Mayo Clinic using over 250 AI models to enhance diagnostic efficiency and accuracy. 🖥️
• Radiology experts describe AI as a "second pair of eyes," capable of handling repetitive tasks, but clinical judgment still relies on doctors. 👀

Related Links:

• AI Godfather Admits Mistake! Geoffrey Hinton Backtracks: I Misjudged, Shouldn’t Have Sentenced Radiologists to a "Slow Death"
• AI Godfather Admits Mistake! Geoffrey Hinton Backtracks: I Misjudged, Shouldn’t Have Sentenced Radiologists to a "Slow Death"

通用提示词策略可攻破主流AI模型安全限制

简报：

• HiddenLayer研究人员发现一种通用提示词策略，可绕过ChatGPT、Claude、Gemini等主流大模型的安全限制
• 该技术通过将危险指令伪装成XML或JSON配置片段，配合无害角色扮演，欺骗模型忽视安全限制
• 攻击策略仅需不足200字符的提示词，无需暴力破解，且对所有主流模型有效
• 该方法利用了模型训练时的系统性弱点，根植于训练数据中，难以通过简单修复解决

相关链接：

• 一个提示攻破所有模型，OpenAI谷歌无一幸免！
• 一个提示攻破所有模型，OpenAI谷歌无一幸免！

Universal Prompt Strategy Bypasses Safety Restrictions of Major AI Models

Newsletter:

• HiddenLayer researchers have uncovered a universal prompt strategy that can bypass safety restrictions on leading AI models like ChatGPT, Claude, and Gemini. 😲
• This technique disguises dangerous instructions as XML or JSON configuration snippets, paired with harmless role-playing, tricking the models into ignoring safety protocols.
• The attack strategy requires prompts of fewer than 200 characters, does not rely on brute force, and is effective against all major models. 💻
• This method exploits systemic weaknesses from the models’ training process, deeply rooted in the training data, making it difficult to address with simple fixes. 🛠️

Related Links:

• One Prompt Breaks All Models, OpenAI and Google Not Spared!
• One Prompt Breaks All Models, OpenAI and Google Not Spared!

美方修改华为芯片限制措辞但仍扩大管制范围

简报：

• 美国商务部工业与安全局5月13日废除《人工智能扩散规则》，但同步出台三项强化全球芯片出口管制新措施
• 新规删除"全球禁用华为Ascend芯片"的绝对化表述，改为强调使用特定型号昇腾芯片可能违反美国出口管制
• 被点名的芯片型号包括昇腾910B、910C及尚未发布的910D，这些7纳米工艺AI加速器被认定含有美国技术元素
• 美国商务部通过援引GP10条款扩大管辖范围，任何主体若知晓涉及受控技术的交易违规仍参与其中可能被追责
• 中国外交部5月16日回应称美方做法严重违反市场规则，扰乱全球产供链，中方坚决反对并将采取坚决措施维护权益

相关链接：

• 美方加严限制华为“昇腾”芯片，外交部：绝不接受！
• 突发！美：全面取消封杀华为芯片！但......
• 涉华为芯片 中方敦促美方停止无底线打压

U.S. Modifies Wording on Huawei Chip Restrictions but Expands Control Scope

Briefing:

• On May 13, the U.S. Department of Commerce's Bureau of Industry and Security revoked the "AI Proliferation Rule," but simultaneously introduced three new measures to strengthen global chip export controls.
• The new regulations remove the absolute phrasing of a "global ban on Huawei Ascend chips," instead emphasizing that using specific Ascend chip models may violate U.S. export controls.
• The named chip models include Ascend 910B, 910C, and the unreleased 910D, which are 7-nanometer AI accelerators identified as containing U.S. technology elements.
• The U.S. Department of Commerce expanded its jurisdiction by invoking the GP10 clause, holding any entity accountable if they knowingly participate in transactions involving controlled technology violations.
• On May 16, China's Ministry of Foreign Affairs responded, stating that the U.S. actions severely violate market rules and disrupt global supply chains, and China firmly opposes this and will take resolute measures to safeguard its rights. 🇨🇳💪

Related Links:

• U.S. Tightens Restrictions on Huawei's "Ascend" Chips, Foreign Ministry: Absolutely Unacceptable!
• Breaking! U.S.: Complete Ban on Huawei Chips Lifted! But...
• Regarding Huawei Chips, China Urges U.S. to Stop Unrestrained Suppression 😠