The Download: OpenAI’s predictable hack, and an AI stock sell
This is today's edition of The Download , our weekday newsletter that provides a daily dose of what's going on in the world of technology. —Will Douglas Heaven, senior AI editor Reading OpenAI’s account last week of how some of its models broke their containment and hacked into the computer systems of Hugging Face, another AI company, was the first time I got genuine chills about what large language models are now able to do. But this is a case of human hubris, not rogue AI. I am not an alarmist. In fact, I have been pushing back against AI scare stories for years . Even so, this incident crossed a line. I think it’s the clearest illustration yet of how the people building and testing this technology do not fully understand what they’re doing. Here’s why OpenAI could—and should—have seen this coming . This st
This is today's edition of The Download , our weekday newsletter that provides a daily dose of what's going on in the world of technology. —Will Douglas Heaven, senior AI editor Reading OpenAI’s account last week of how some of its models broke their containment and hacked into the computer systems of Hugging Face, another AI company, was the first time I got genuine chills about what large language models are now able to do. But this is a case of human hubris, not rogue AI. I am not an alarmist. In fact, I have been pushing back against AI scare stories for years . Even so, this incident crossed a line. I think it’s the clearest illustration yet of how the people building and testing this technology do not fully understand what they’re doing. Here’s why OpenAI could—and should—have seen this coming . This st
ory is from The Algorithm, our weekly AI newsletter. Sign up to receive it in your inbox every Monday. I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 There’s a growing global AI stock sell-off underway 📉 Chip and memory stocks are bearing the brunt so far. ( FT $) + It was partly sparked by a report that a Chinese company has started making a key piece of chip equipment for the first time. ( The Information $) + Like their US rivals, Chinese AI firms are struggling to find a path to profitability. ( NYT $) + What even is the AI bubble? ( MIT Technology Review ) 2 Some people’s chats with Claude were open to anyone online OpenAI had a near-identical issue with ChatGPT last year. ( BBC ) + Is a secure AI assistant even possible? ( MIT Technology Review ) 3 France just witnessed its first ever “fire cloud” It’s a scary sign of the intensity of the blazes burning there. ( Wired $) 4 Meta is screwing up its smart glasses rollout Privacy issues keep cropping up, and its response is invariably too little, and too late. ( The Verge $) + They even have a nickname: “pervert glasses.” ( Vox $) 5 Efforts are underway to measure Spotify’s AI slop problem People are desperate for it to start labelling AI-generated music. ( 404 Media ) + We’re in a weird era regarding AI’s role in literature. ( The Atlantic $) 6 People are renting out their faces to AI in China Often for microdramas, which are big business and increasingly rely on AI. ( Rest of World ) + How Chinese short dramas became AI content machines. ( MIT Technology Review ) 7 Microsoft is having a torrid year Rivals building better AI tools are threatening its business on multiple fronts. ( Business Insider $) 8 How bots took over the internet They now outnumber humans, in terms of overall web traffic. ( New Yorker $) 9 Robotaxis are hitting London’s streets this summer Just for testing for now, though there are plans to launch rides to the public next year. ( Engadget ) 10 Why video games are obsessed with Backrooms These spaces, which players aren’t supposed to access, can provoke a strange, liminal sort of discomfort. ( Guardian ) —Christine Peterson, co-founder of the grantmaking group Foresight Institute, tells Wired charities anticipate a huge philanthropy windfall from Anthropic and OpenAI IPOs. On the ground in Ukraine’s largest Starlink repair shop Starlink is critical to Ukraine’s ability to continue in the fight against Russia, but Donald Trump’s fickle foreign policy and reports suggesting Elon Musk might remove Ukraine’s access to the services have cast the technology’s future in the country into doubt. For now Starlink access largely comes down to the unofficial community of users and engineers, including the expert “Dr. Starlink”—famous for his creative ways of customizing the systems—who have kept Ukraine in the fight, both on and off the front line. He gave MIT Technology Review exclusive access to his unofficial Starlink repair workshop in the city of Lviv. Read the full story . We can still have nice things A place for comfort, fun, and distraction to brighten up your day. (Got any ideas? Drop me a line .) + As these retirees have discovered, dancing is incredibly good for you. + Do you feel lucky, punk? Life feels a lot easier if you do! + 100% would visit this cat every day if I lived in Istanbul. + Watch this to understand how the movie Sinners had such impressive visual effects. Plus: OpenAI has unveiled its
long-awaited "super app." Plus: New York has become the first state to enact a data center moratorium. Plus: Meta is pausing an AI training program that tracks workers’ keystrokes. Plus: Anthropic has called for a global slowdown in AI development. Discover special offers, top stories, upcoming events, and more. Privacy Policy Thank you for submitting your email! It looks like something went wrong. We’re having trouble saving your preferences. Try refreshing this page and updating them one more time. If you continue to get this message, reach out to us at customer-service@ with a list of newsletters you’d like to receive.
反向传播详解（二）
在上一部分中，我们通过一个小型神经网络理解了前向传播过程，发现预测值与真实值相差较大，需要训练网络。损失函数依赖于七个参数，我们开始对w1进行经典微分。然而，面对剩余的六个参数，是否要用相同方法？本文引入链式法则来解决这一问题。通过分析w1的变化如何经z1、激活值a1、预测值ŷ最终影响损失L，我们注意到神经网络中的依赖链条更长，但原理与简单例子一致。链式法则允许我们将复杂求导分解为沿着依赖路径的多个小步骤，从而高效计算每个参数的梯度。这一思想构成了反向传播算法的核心，使得训练多层网络成为可能。 #反向传播 #链式法则 #神经网络 #梯度下降 #机器学习 #AI教程 #数学原理
在上一部分中，我们通过一个小型神经网络理解了前向传播过程，发现预测值与真实值相差较大，需要训练网络。损失函数依赖于七个参数，我们开始对w1进行经典微分。然而，面对剩余的六个参数，是否要用相同方法？本文引入链式法则来解决这一问题。通过分析w1的变化如何经z1、激活值a1、预测值ŷ最终影响损失L，我们注意到神经网络中的依赖链条更长，但原理与简单例子一致。链式法则允许我们将复杂求导分解为沿着依赖路径的多个小步骤，从而高效计算每个参数的梯度。这一思想构成了反向传播算法的核心，使得训练多层网络成为可能。 #反向传播 #链式法则 #神经网络 #梯度下降 #机器学习 #AI教程 #数学原理
AI Agent能否真做科研？Agentic Science Challenge用万道题验证
DeepModeling社区发起Agentic Science Challenge，旨在检验AI Agent独立完成科学任务的能力。竞赛提供超过1万道从科学文献中提取的可验证题目，覆盖生物、物理、化学、材料等领域，并按难度分级。参赛系统需完成问题理解、方案规划、工具调用、代码执行与结果验证等全流程。平台支持170多种编程语言和5万多种科学工具，并开放基础算力。评分采用结果分与轨迹分双维度加权，既关注最终答案也重视思考过程。比赛自7月31日启动轮赛，每周一轮，优胜者可获Mac mini等奖品。赛事强调解题路径的沉淀与社区共享，旨在探索AI Agent在科研中的实践前景。 #AI #Agent #智能体 #科学竞赛 #AIforScience #DeepModeling #科研 #科技新闻
DeepModeling社区发起Agentic Science Challenge，旨在检验AI Agent独立完成科学任务的能力。竞赛提供超过1万道从科学文献中提取的可验证题目，覆盖生物、物理、化学、材料等领域，并按难度分级。参赛系统需完成问题理解、方案规划、工具调用、代码执行与结果验证等全流程。平台支持170多种编程语言和5万多种科学工具，并开放基础算力。评分采用结果分与轨迹分双维度加权，既关注最终答案也重视思考过程。比赛自7月31日启动轮赛，每周一轮，优胜者可获Mac mini等奖品。赛事强调解题路径的沉淀与社区共享，旨在探索AI Agent在科研中的实践前景。 #AI #Agent #智能体 #科学竞赛 #AIforScience #DeepModeling #科研 #科技新闻
XYZ AI Lab 发布 Aquila 系列，用 AI4AI 方法让 AI 自主改进 AI
XYZ AI Lab 近日正式发布两款面向深度搜索场景的 Agent：Aquila-mini 和 Aquila-pro，在多项基准测试中取得领先成绩。然而，更值得关注的是其技术报告披露的 AI4AI 研发方法——AI 不仅是被训练出的产品，还开始参与问题发现、方案提出和验证改进。在研发循环中，AI Agent 读取评测结果和失败轨迹，自主分析问题根源（数据、训练、工具或上下文），提出候选改进并落实为代码或配置，再通过实验验证有效性。团队在接受采访时强调，AI4AI 的关键在于 AI 是否进入研发判断闭环，而非简单的自动化辅助。与 RSI（递归式自我改进）不同，Aquila 目前实现的是有边界的系统级递归改进：人类仍控制目标、评测和权限，AI 在预设范围内迭代优化系统组件。这一范式有望改变 Agent 的研发方式，推动 AI 从工具向协作者演进。 #AI #DeepSearch #AI4AI #RSI #Agent #XYZ #Aquila #大模型 #科技新闻
XYZ AI Lab 近日正式发布两款面向深度搜索场景的 Agent：Aquila-mini 和 Aquila-pro，在多项基准测试中取得领先成绩。然而，更值得关注的是其技术报告披露的 AI4AI 研发方法——AI 不仅是被训练出的产品，还开始参与问题发现、方案提出和验证改进。在研发循环中，AI Agent 读取评测结果和失败轨迹，自主分析问题根源（数据、训练、工具或上下文），提出候选改进并落实为代码或配置，再通过实验验证有效性。团队在接受采访时强调，AI4AI 的关键在于 AI 是否进入研发判断闭环，而非简单的自动化辅助。与 RSI（递归式自我改进）不同，Aquila 目前实现的是有边界的系统级递归改进：人类仍控制目标、评测和权限，AI 在预设范围内迭代优化系统组件。这一范式有望改变 Agent 的研发方式，推动 AI 从工具向协作者演进。 #AI #DeepSearch #AI4AI #RSI #Agent #XYZ #Aquila #大模型 #科技新闻
腾讯加速布局AI全产业链，从大模型到芯片深度落子
腾讯正将投资重心从传统互联网转向人工智能，全面布局大模型、AI Agent和芯片赛道。近期减持快手股份，同时参与其旗下Kling AI的30亿美元融资。在大模型领域，腾讯投资了月之暗面（Moonshot AI）、深度求索（DeepSeek）、智谱AI和MiniMax，其中对DeepSeek的A轮投资约100亿元。AI Agent公司Manus持续获腾讯支持，其考虑增持成为最大外部股东。在AI芯片方面，腾讯持有燧原科技20.26%股份，是其最大客户，贡献其84%营收；还投资了光子芯片公司Lightelligence。业内认为，腾讯正将生成式AI、大模型、Agent与半导体整合为清晰投资主线，以巩固其在AI产业中的战略位置。 #腾讯 #AI #大模型 #AI芯片 #投资 #月之暗面 #深度求索 #燧原科技 #科技新闻 #企业动态
腾讯正将投资重心从传统互联网转向人工智能，全面布局大模型、AI Agent和芯片赛道。近期减持快手股份，同时参与其旗下Kling AI的30亿美元融资。在大模型领域，腾讯投资了月之暗面（Moonshot AI）、深度求索（DeepSeek）、智谱AI和MiniMax，其中对DeepSeek的A轮投资约100亿元。AI Agent公司Manus持续获腾讯支持，其考虑增持成为最大外部股东。在AI芯片方面，腾讯持有燧原科技20.26%股份，是其最大客户，贡献其84%营收；还投资了光子芯片公司Lightelligence。业内认为，腾讯正将生成式AI、大模型、Agent与半导体整合为清晰投资主线，以巩固其在AI产业中的战略位置。 #腾讯 #AI #大模型 #AI芯片 #投资 #月之暗面 #深度求索 #燧原科技 #科技新闻 #企业动态
Moonshot AI 声称 Kimi K3 是全球首个开源 3T 级大模型
据一财全球报道，人工智能初创公司 Moonshot AI 宣布其最新大模型 Kimi K3 已正式开源，并声称这是全球首个参数规模达到 3 万亿（3T）级别的开源大模型。这一里程碑标志着开源社区首次获得如此大规模的语言模型，意味着科研人员和开发者可以直接使用或基于该模型进行微调与二次开发，有望推动多模态理解、长上下文推理等前沿领域的研究。不过，Moonshot AI 目前尚未披露 Kimi K3 的具体训练细节及评测结果，业界对其实际性能仍持观望态度。 #MoonshotAI #KimiK3 #开源 #大模型 #AI #3T模型 #科技新闻
据一财全球报道，人工智能初创公司 Moonshot AI 宣布其最新大模型 Kimi K3 已正式开源，并声称这是全球首个参数规模达到 3 万亿（3T）级别的开源大模型。这一里程碑标志着开源社区首次获得如此大规模的语言模型，意味着科研人员和开发者可以直接使用或基于该模型进行微调与二次开发，有望推动多模态理解、长上下文推理等前沿领域的研究。不过，Moonshot AI 目前尚未披露 Kimi K3 的具体训练细节及评测结果，业界对其实际性能仍持观望态度。 #MoonshotAI #KimiK3 #开源 #大模型 #AI #3T模型 #科技新闻