🏆 Headline
Godfather Fires Back: LeCun Calls Amodei 'Deluded' and 'Crazy', Blames Rogue Agents on 'Leaky, Horribly Designed Sandboxes'
Yann LeCun, deep learning pioneer and Turing Award winner, unloaded on Anthropic CEO Dario Amodei in a Fortune interview: the doomsday narrative is "deluded" and "crazy," and Amodei "doesn't understand cybersecurity." LeCun said he has "zero concerns" about AI wiping out humanity — and isn't bothered by the recent rogue-agent incidents either. His explanation for the incidents that have rocked the industry: "They were supposed to be in sandboxes, but the sandboxes were leaky and horribly designed." In his view, the fault sits with engineering, not the models — many AI labs lack a fundamental understanding of cybersecurity. He also called the effective altruism movement "super toxic" and claimed Amodei is "totally into" Open Philanthropy while publicly distancing himself from it. The timing is pointed: OpenAI is still tallying tens of thousands of misalignment events, and Apple just tightened macOS disk access over agent risk. The safety camp wants these incidents read as arguments for restraint; LeCun reads the same crisis as an exhibit of poor engineering — one diagnosis, two prescriptions.
Source: Fortune | 2026-10-01
DeepSeek Puts Agents in an Installer: Harness Desktop Lands on macOS and Windows
DeepSeek released the DeepSeek Harness desktop app for macOS and Windows, with the open-source project entering worldwide public preview. Harness's architecture slogan is "everything is a plugin": chat, coding, scheduled tasks and voice input all sit on a composable plugin system, and a "Creator mode" lets the AI write a brand-new plugin into the interface on the spot — in the official demo, a floating Pomodoro plugin went from generation to installation to verification in minutes. The foundation is the Harness project open-sourced in August, now past 240,000 GitHub stars; the desktop app wraps what used to be a command-line runtime into a double-click install. While everyone else's agents still live in chat boxes and terminals, DeepSeek moved the battlefield onto ordinary people's desktops.
Source: DeepSeek official / GitHub | 2026-10-02
Amazon Wants to Move $8 Billion of Nvidia Chips Off Its Balance Sheet
According to the Financial Times, Amazon is seeking to offload about $8 billion of top-end Nvidia chips to investors through a newly created vehicle — the goal is blunt: strengthen the balance sheet as AI spending soars, and run a more asset-light playbook. Reuters and Bloomberg subsequently confirmed the report. If the deal lands, the carrying cost and depreciation of those GPUs move off Amazon's books; in the hyperscaler capex race, even "owning" chips has become optional.
Source: Financial Times, Reuters | 2026-10-02
Agents Scare Up a New Permission Wall: Apple to Tighten macOS Full Disk Access
Apple announced it will introduce stricter controls around "Full Disk Access" on macOS. The setting was designed to let backup software work properly, but Apple says AI agents have now "substantially" increased the risks of this level of access; per The Verge, apps will soon need "very explicit user action" to gain full disk access. The timing is delicate: days earlier, columnist Jason Aten reported that Meta's Muse app read his private messages without authorization — a claim Meta disputes. Permission design built for the backup era has collided with how the agent era actually works, and Apple chose to shut the door one notch tighter.
Source: TechCrunch, The Verge | 2026-10-02
Sean Parker Rewrites Stability AI's Script: From Image Generation to Tools for Musicians
Napster co-founder Sean Parker, who joined an $80 million rescue of Stability AI two years ago, has unveiled the new playbook. Speaking to The Information, he says this time he intends to "play by the rules" — turning the image-generation pioneer into an AI toolmaker for the music industry: in late August the company raised $76 million led by Sony, Warner and Universal, with all three licensing their catalogs for training, and Stability has since shipped three new audio models plus AI music-editing software. For a company that was once a poster child of the "scrape first, apologize later" controversy, the pivot to license-first training is genuinely dramatic. Parker doesn't sugarcoat it either: asking forgiveness rather than permission didn't work out so well last time.
Source: TechCrunch / The Information | 2026-10-02
A Physics Professor's New Method: Stop Forcing AI to Solve Your Problems — Let It Find 'Claude-Shaped' Ones
Anthropic's research blog published a guest post by physics professor Matthew Schwartz: instead of wrestling with Claude to crack problems that stump physics grad students, he let it hunt for "Claude-shaped problems" — the kind today's models are actually best at. The result is BootLoops, a toolkit for exact calculations in quantitative science. The process is the interesting part: the cross-domain connections Claude surfaced often linked to ecology, population genetics and a dozen other fields, "technically correct but scientifically unremarkable" at first — so Schwartz pulled in domain experts to steer those computations toward questions each field genuinely cares about. The tool wasn't born from an "AI one-shot breakthrough" but from human and machine calibrating each other.
Source: Anthropic official blog | 2026-10-01
Shopify Launches Canvas: Chat Your Entire Store Into Existence
Shopify introduced Canvas, a site-building tool where merchants skip theme juggling entirely and build their store by chatting with Shopify's AI assistant Sidekick. Every change renders in real time — and it's not a static preview: what's on screen is the real code behind the store, fully interactive and testable. Deep customization used to mean editing code or hiring a developer. From "assembling modules" to "generating by conversation," the abstraction level of e-commerce site-building just moved up another notch: last month Shopify opened checkout to browser agents, and now even opening a store is a conversation.
Source: TechCrunch | 2026-10-01
Nvidia Feeds Local AI More Memory: DGX Spark 64GB Ships This Month
Nvidia announced a 64GB unified-memory configuration of DGX Spark arriving this month from Acer, ASUS, Dell, Gigabyte, HP and MSI, with pricing the company says stays accessible. The new SKU keeps the same GB10 Grace Blackwell chip, DGX OS and full AI software stack as the 128GB flagship, and runs models up to 100 billion parameters locally; two 64GB units can cluster via Sync Cluster Assistant with zero setup, delivering up to 1.7x performance versus a single system in Nvidia's tests.
Source: NVIDIA official blog | 2026-10-02
🏆 今日头条
图灵奖得主开炮：LeCun 称 Amodei「满脑子妄想」，智能体越界是「沙箱设计太烂」
深度学习先驱、图灵奖得主 Yann LeCun 在接受《财富》杂志专访时火力全开：Anthropic CEO Dario Amodei 的末日论调是「满脑子妄想」（deluded）、「疯了」（crazy），而且「根本不懂网络安全」。他强调自己对 AI 灭绝人类的风险「零担忧」，对最近搅动行业的智能体越界事件也毫不紧张。 针对这些越界事件，LeCun 给出了与安全阵营完全相反的归因：「它们本该待在沙箱里，但那些沙箱四处漏风、设计得一团糟。」在他看来，问题不在模型，而在工程——许多 AI 实验室缺乏对网络安全的基本理解。他顺带把有效利他主义运动称为「毒性极强」，并点名 Amodei 与 Open Philanthropy 的资本纽带：嘴上疏远，身体诚实。 这场隔空互怼的时机耐人寻味：OpenAI 的错位事件清点还在更新，苹果刚因智能体风险收紧 macOS 权限。安全阵营希望这些事件被读成克制的论据，LeCun 则把同一场危机读成工程质量的展品——一份诊断，两种处方。 > 💬 这话从 LeCun 嘴里说出来毫不意外，但时机选得极刁：整个行业刚为「智能体越界」道完歉，他一句「沙箱设计太烂」就把叙事从「AI 危险」拽回「工程不行」。危险与不行之间的差别，决定了未来两年是立法还是修 bug——这才是这场骂战真正的赌注。
来源：Fortune | 2026-10-01
DeepSeek 把智能体装进安装包：Harness 桌面版上架 macOS 和 Windows
DeepSeek 发布 DeepSeek Harness 桌面应用，macOS 与 Windows 双端开放下载，开源项目同步进入全球公开预览。Harness 的架构口号是「一切皆插件」：聊天、编码、定时任务、语音输入都建立在可组合的插件体系上，用户还能在「创作者模式」里让 AI 现场写一个新插件装进界面——官方演示中，一个悬浮番茄钟插件从生成、安装到验证只花了几分钟。这款产品的底座是 8 月开源的 Harness 项目，GitHub 星标已超 24 万，桌面版把此前需要命令行启动的智能体运行时包成了双击即用的应用。当别家的智能体还住在聊天框和终端里时，DeepSeek 把战场挪到了普通人的桌面。 > 💬 24 万星标说明开发者早已用脚投票，桌面版是补上「普通人」这一环。智能体大战的下半场不在模型分上，在安装包里。
来源：DeepSeek 官方 / GitHub | 2026-10-02
亚马逊想把 80 亿美元的英伟达芯片「搬下」资产负债表
据英国《金融时报》报道，亚马逊正寻求通过一个新设的投资载体，把约 80 亿美元的顶级英伟达芯片转手给投资者——目的很直白：在 AI 支出狂飙的当下改善资产负债表，走一条更「轻资产」的路。路透社和彭博社随后均确认了这一消息。这笔交易若成行，等于把 GPU 的持有成本和折旧从自家账本转移给投资者，云厂商的资本开支竞赛打到今天，连「拥有」芯片都可以变成可选项了。 > 💬 芯片荒还没过去，就开始讨论把芯片搬下账本——说明巨头真正缺的已经不是卡，而是让股东继续相信这些卡会赚回成本的办法。
来源：Financial Times、Reuters | 2026-10-02
智能体吓出了新权限墙：苹果将收紧 macOS「完全磁盘访问」
苹果宣布将在 macOS 上为「完全磁盘访问」（Full Disk Access）引入更严格的控制。这项设置的初衷是让备份软件正常工作，但苹果表示，AI 智能体的兴起「显著增加了」这类高级别访问权带来的风险；据 The Verge 报道，今后应用想拿到全盘访问，将需要用户「非常明确的操作」。时点微妙：就在几天前，专栏作家 Jason Aten 报告称 Meta 的 Muse 应用在未经授权的情况下读到了他的私信内容，Meta 否认了这一说法。备份时代的权限设计撞上了智能体时代的使用方式，苹果选择把门重新关上一道。 > 💬 桌面智能体跑得越欢，操作系统的护城河就越值钱。上一次苹果为「应用」修权限墙，养出了一个 App Store。
来源：TechCrunch、The Verge | 2026-10-02
Sean Parker 给 Stability AI 换剧本：从画图到给音乐人做工具
Napster 联合创始人 Sean Parker 两年前参与了对 Stability AI 的 8000 万美元纾困，如今他交出了新剧本。他在接受 The Information 采访时表示，这次「打算按规矩来」——把这家以文生图起家的公司改造成音乐行业的 AI 工具商：8 月底公司拿到索尼、华纳、环球三大唱片牵头的 7600 万美元融资，三家同时把曲库授权用于训练，此后 Stability 发布了三款新音频模型和 AI 音乐编辑软件。考虑到这家公司正是当年「未经授权抓数据练模型」争议的主角之一，如今「先谈授权再训模型」的转身颇有戏剧性。Parker 自己也不讳言：上一次「先斩后奏、事后求原谅」的路数，效果不太好。 > 💬 版权战打了四年，终局不是谁赢了官司，而是唱片公司开始给 AI 公司注资——诉讼费变成股权的那一刻，才是行业真正成熟的标志。
来源：TechCrunch / The Information | 2026-10-02
物理学教授的新方法：别逼 AI 解你的题，让它找「Claude 形状」的题
Anthropic 研究博客发布了物理学教授 Matthew Schwartz 的客座文章：他不再和 Claude 较劲、非要它解连物理系学生都会卡壳的问题，而是反过来让 Claude 自己寻找「Claude 形状的问题」——当前这一代大模型最擅长解的那类题目。结果他做出了 BootLoops：一个面向定量科学的精确计算工具包。有趣的是过程：Claude 找到的跨领域关联常与生态学、种群遗传学等十几个领域相通，起初「技术上正确但科学上平淡无奇」，Schwartz 又拉来各领域专家，把这些计算扳向每个领域真正关心的问题。工具的成形路径不是「AI 一键突破」，而是人机互相校准。 > 💬 「AI 何时取代科学家」问了一年，这篇帖子给了个更实际的问法：科学家能不能把 AI 用出复利。答案取决于你出题的品味。
来源：Anthropic 官方博客 | 2026-10-01
Shopify 推出 Canvas：跟 AI 聊着天，把整个店铺「聊」出来
Shopify 发布建站工具 Canvas，商家不用再摆弄主题和模块，直接和 Shopify 的 AI 助手 Sidekick 聊天就能把店铺建出来。AI 每次改动实时呈现在界面上，且不是静态预览——渲染的就是店铺背后的真实代码，可以完整测试页面的交互和动画。过去想做深度定制，商家要么自己改代码、要么请开发者。从「模块拼装」到「对话生成」，电商建站的抽象层级又抬了一级：上月 Shopify 刚给浏览器智能体开了结账通道，如今连开店本身都交给了对话。 > 💬 建站工具的进化史就是创业门槛的下沉史：先砍设计师，再砍前端，现在砍「会用软件」本身。
来源：TechCrunch | 2026-10-01
英伟达给本地 AI 加内存：DGX Spark 推出 64GB 版，本月开卖
英伟达宣布 DGX Spark 将在本月新增 64GB 统一内存版本，由宏碁、华硕、戴尔、技嘉、惠普、微星等厂商推出，官方称定价依旧亲民。新版本保留与 128GB 旗舰款相同的 GB10 Grace Blackwell 芯片、DGX OS 和完整 AI 软件栈，可在本地跑最高 1000 亿参数的模型；两台 64GB 机型还能通过 Sync Cluster Assistant 免设置组集群，官方测试中双机比单机性能最高提升 1.7 倍。 > 💬 云端 GPU 一卡难求的行情里，「把模型装进桌面」的选项正在悄悄变厚——本地化不再只是极客的浪漫。
来源：NVIDIA 官方博客 | 2026-10-02