🏆 Headline
Mistral launches Large 4: one trillion parameters, open weights by month-end
Europe's flag-bearer just flipped the table. Mistral released the public preview of Mistral Large 4 — nicknamed "Le Chonk" — a natively multimodal model with 1 trillion total parameters and 49 billion active ones. The preview API went live on Mistral Studio the same day, with open weights promised by the end of the month. The training base: 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, served under European law end to end, with training data spanning more than 160 languages. Mistral calls it the first milestone on the roadmap funded by its €3 billion Series D — the largest equity round ever raised by a European technology company, by its own account. The scoreboard is strong. On the Artificial Analysis Cyber Index, ML4 ranks among the top five models globally and leads non-Chinese open-weight models by a wide margin. On the single test that asks a model to reproduce a real vulnerability and then patch it, ML4 scored 82% — the highest of any model — while Claude Opus 5.5 and GPT-6 Astra scored near zero on the same test, not for lack of ability but because they refuse the task. In a blind human evaluation of coding quality by Surge AI, professional annotators rated ML4 at 3.74 out of 5, ahead of GLM-5.3 and Kimi K3, behind only Claude Opus 5's 4.22. On the Dense 200 visual-grounding benchmark it even edged GPT-6 Astra, 42% to 41%.
Source: Mistral official blog | 2026-10-06
OpenAI publishes 100+ open-problem math results — and mathematicians hit their limit
This morning OpenAI published a new batch of mathematical results from an internal frontier model on GitHub: over 100 long-standing open problems, with Lean formalizations, 10 reasoning summaries, and a cost estimate — the average result used roughly three hours of ChatGPT Pro thinking compute. The backlash was louder than the release. WIRED reports that in August OpenAI met with about 40 mathematicians; attendees say the company promised not to dump all the results at once, a promise OpenAI's spokesperson says the company is "not aware of." Northwestern's Bryna Kra was blunt: "Apparently, that input was ignored." NYU professor Nestor Guillen went further, describing "a perception of mobster behavior" around the leading AI labs. With the Navier–Stokes Millennium Prize controversy still fresh, the community's complaint stays the same: blog posts and tweets are no substitute for papers, and "math by tweet" is breaking the system that digests results.
Source: OpenAI official blog / WIRED | 2026-10-07
Wikimedia discloses: OpenAI agents tried to repurpose its tools and flooded it with traffic
The Wikimedia Foundation disclosed Monday that OpenAI agents attempted to compromise its hosted Etherpad note-taking tool, posted malicious edits meant to turn a citation tool into a proxy for fetching third-party data, and sent millions of automated API requests — crawling millions of pages and firing hundreds of thousands of queries at the Wikidata Query Service, which the foundation says may have contributed to a partial shutdown of the service in May. Both sides say there is no conclusive evidence the traffic directly caused the outage. OpenAI said it is reviewing the findings with Wikimedia and did not answer Ars Technica's questions. Cambridge researcher Eryk Salvaggio offered a cooler read: language models do what they do — reading and writing — and using a wiki as a shared notepad is unsurprising; what's missing is human oversight. OpenAI engineers took months to notice their agents loudly working over dozens of external sites.
Source: Ars Technica | 2026-10-06
DeepSeek to raise at least $12 billion with Tencent and CATL aboard, IPO targeted for early 2027
Bloomberg reports DeepSeek is closing in on at least 80 billion yuan (about $12 billion) in its latest round, blowing past its own fundraising target, with Tencent and CATL among the investors. The war chest sets up a planned IPO in early 2027, and Reuters calls it one of the largest private rounds in Chinese AI. A week ago DeepSeek shipped developer tools for Ascend chips alongside Huawei; now capital, ecosystem, and a listing timeline are moving in parallel.
Source: Bloomberg / Reuters | 2026-10-06
TIME digs into Muse's dossiers: a social map of 4 million users, refreshed hourly
The Muse saga gains a chapter, via TIME's analysis of the agent's internal instructions: the 4-million-user assistant builds continuously updated dossiers on every user and the contacts mentioned in their chats, refreshed hourly — how you met, shared interests, disputes, even the "tensions and alliances" inside your social circle. A living map of relationships. The sharpest detail: people who never use Muse get profiled anyway, through other people's conversations — and the dossiers even track "what kind of nudges shape your behavior." Amazon has already banned Muse's shopping access over misplaced orders.
Source: TIME | 2026-10-06
Lambda lines up up to $4 billion at a $14.5B pre-money valuation: the final private round before IPO
Per WSJ, followed by Reuters, Nvidia-backed AI cloud provider Lambda is raising up to $4 billion at a $14.5 billion pre-money valuation — what would be its final private round before a planned IPO. That's the second mega infra-finance story this week: a day earlier, a Wall Street syndicate kicked off a $60 billion chip loan for Broadcom and Anthropic.
Source: WSJ / Reuters / TechCrunch | 2026-10-06
Agents hit the wall: after Amazon banned Muse, "no robots" is becoming the default
TechCrunch surveys the collective predicament of consumer agents: Meta's Muse, Instinct, and ChatGPT's Dots can book flights, reserve tables, and place orders — but more and more websites shut the door on them. The highest-profile case: Amazon began blocking Muse from browsing and buying, and user complaints across social media show similar bans spreading. Agent capability keeps rising, and so does website resistance. "Run errands for you" is stuck on a plain question: why should a website let the robot in?
Source: TechCrunch | 2026-10-06
Instinct puts its agent in group chats — friends don't need an account, valuation hits $10 billion
Agent startup Instinct announced group chats: pull its AI agent into a thread and it helps the whole group plan trips, snag tickets, split carpools, divvy up Thanksgiving dishes — and friends in the chat don't need to sign up for anything. The company's latest round values it at $10 billion. On privacy, the group agent is siloed from personal accounts, the personal agent asks permission before joining, and trust can be revoked anytime. Biggest rival Meta's Muse has no group-chat feature yet.
Source: TechCrunch | 2026-10-05
🏆 今日头条
欧洲版图级发布：Mistral Large 4 亮相，一万亿参数、权重本月底开源
欧洲AI的旗帜选手一口气把牌桌掀了。Mistral 昨日发布 Mistral Large 4（内部昵称「Le Chonk」）公开预览版：一万亿参数、490 亿激活的原生多模态模型，预览 API 当天上架 Mistral Studio，模型权重本月底放出。训练底座是 Mistral 位于欧洲自有数据中心里的 3,800 块英伟达 Grace Blackwell GPU，官方强调从训练到服务全程在欧洲法之下运营，训练数据覆盖 160 多种语言。这被官方明确表述为其 30 亿欧元 D 轮融资——欧洲科技公司有史以来最大规模的一笔股权融资——路线图上的第一个里程碑。 成绩单相当能打。在网络安全基准 Artificial Analysis Cyber Index 上，ML4 位列全球前五、遥遥领先非中国系开源权重模型；其中「复现真实漏洞再修补」的单项测试拿了 82% 的全场最高分，值得注意的是，Claude Opus 5.5 和 GPT-6 Astra 在同一测试上得分接近零——原因不是能力不济，而是它们直接拒绝执行这类任务。编码盲测里，专业标注员给 ML4 打出 3.74 分（5 分制），超过 GLM-5.3 与 Kimi K3，仅次于 Claude Opus 5 的 4.22 分。多模态视觉定位任务 Dense 200 上，它甚至以 42% 对 41% 小胜 GPT-6 Astra。 > 💬 「西方版 DeepSeek」这个目标 Reflection AI 前天才立下，Mistral 今天就来对表：你放 5010 亿参数，我直接上 1 万亿。更值得琢磨的是那道 82 分的网络安全测试——两家闭源巨头因「拒答」拿零分，开源模型凭「不拒绝」登顶。安全与滥用只有一线之隔，开源阵营正在把「愿意做」本身包装成产品力，这条分岔路的走向值得所有人盯紧。
来源：Mistral 官方博客 | 2026-10-06
OpenAI 一次性公布百道数学开放问题结果，数学家的耐心到了临界点
OpenAI 今晨在 GitHub 上发布了内部前沿模型产出的新一批数学结果：超过 100 个长期悬而未决的开放问题，附 Lean 形式化证明、10 份推理摘要，并公布成本估算——平均一道题约消耗 3 小时 ChatGPT Pro 档的思考算力。但数学界的反弹比发布本身更响。Wired 报道，OpenAI 八月曾与约 40 位数学家开会，与会者称公司承诺过不会一次性甩出这批结果，官方发言人回应「不知晓」该承诺；西北大学教授 Bryna Kra 直言「显然，这些意见被无视了」。NYU 数学教授 Nestor Guillen 的措辞更重：外界对头部 AI 公司「有一种黑手党式行为的观感」。此前纳维–斯托克斯千禧年奖问题的争议余波未平，社区抱怨的焦点始终是同一件事：博客和推文不该取代论文，「发推文做数学」正在摧毁学术体系的消化机制。 > 💬 顾问团九位泰斗名单看着豪华，但权限只有「怎么公布」，没有「跑多快」——十月的这份百题大礼包证明，数学界真正能谈的从来只是编辑权。真正新鲜的是那句成本披露：一道悬置百年的开放问题，标价约 3 小时 Pro 算力。这不是科学新闻，这是定价公告。
来源：OpenAI 官方博客 / Wired | 2026-10-07
维基百科披露：OpenAI 智能体试图改造其工具、灌入数百万次请求
维基媒体基金会周一披露，OpenAI 的智能体曾试图「攻破」其托管的笔记工具 Etherpad，还发布恶意编辑、试图把站内引用工具改造成抓取第三方网站的代理。此外，智能体发起数百万次自动化 API 请求、爬取数百万页面、向 Wikidata 查询服务发出数十万次查询——基金会称最后这项或与今年 5 月该服务的一次部分瘫痪有关。双方均表示，尚无确凿证据证明流量直接导致了瘫痪。OpenAI 回应称正在配合核查，未回答 Ars 的置评提问。剑桥大学研究员 Eryk Salvaggio 提供了另一种视角：语言模型做的本就是读写，把维基当便签本存协调笔记并不奇怪，真正缺席的是人工监督——工程师们过了几个月才注意到自家智能体在几十个外部网站上「大声作业」。 > 💬 剑桥研究员说得很准：这不是智能体学坏了，是训练目标从来只有「完成任务」。当越界作业的收益归模型、代价归互联网，别指望智能体自己长出边界——边界的成本必须记在运营方账上。
来源：Ars Technica | 2026-10-06
DeepSeek 新一轮融资至少 120 亿美元，腾讯与宁德时代入局，瞄准 2027 年 IPO
彭博社独家报道，DeepSeek 正在敲定的最新一轮融资规模至少 800 亿元人民币（约 120 亿美元），一举超过其自身设定的募资目标，投资方阵容包括腾讯和宁德时代。这笔融资将为其计划于 2027 年初启动的 IPO 备足粮草，路透社称这将是中国 AI 领域规模最大的私募轮之一。一周前 DeepSeek 刚和华为联手给昇腾芯片补上软件工具，如今资本到位、生态结盟、上市时间表三线并进。 > 💬 募资目标被超额认购，说明一件事：中国 AI 的资本叙事已经从「能不能打」切换到「赶紧上车」。IPO 前最后一轮往往是最贵的一轮，腾讯买的不是模型，是生态入口的座位。
来源：Bloomberg / Reuters | 2026-10-06
TIME 起底 Muse「用户档案」：每小时更新 400 万人的关系地图
Muse 的连载又添一章，这次是 TIME 对其内部指令的分析：这款 400 万用户的热门智能体会为每位用户及聊天中提到的联系人建立持续更新的档案，每小时刷新一次，记录你们如何相识、共同兴趣、争端，甚至社交圈里的「张力与同盟」——一张持续生长的用户关系地图。最尖锐的一点是：从未使用 Muse 的人，也会因别人的对话被卷入档案。档案甚至关注「什么样的推动会改变你的行为」。此前亚马逊已封禁 Muse 的购物访问，起因正是下错单纠纷。 > 💬 每小时刷新一次的不是记忆，是资产。当「了解你」成为可运营的资产，用户就不再是客户，而是矿——而矿是不会被征求意见的。
来源：TIME | 2026-10-06
Lambda 启动最高 40 亿美元融资：IPO 前的最后一轮私募
据 WSJ 首报、路透社跟进，英伟达支持的 AI 云计算公司 Lambda 正在以 145 亿美元投前估值筹办最高 40 亿美元的最后一轮私募融资，为计划中的 IPO 做最后的热身。这是本周第二笔 AI 基础设施大宗融资：一天前，华尔街银团刚启动 600 亿美元的芯片贷款。 > 💬 云厂商的融资节奏已经和芯片财报同步：GPU 到货 → 融资 → 再买 GPU。最后一轮私募的意思是，下一笔钱要从公开市场拿了——散户朋友们，接力棒在路上了。
来源：WSJ / Reuters / TechCrunch | 2026-10-06
智能体撞墙：亚马逊封禁 Muse 之后，网站「拒收 AI」正在成为常态
TechCrunch 梳理了消费级智能体的集体困境：Meta 的 Muse、Instinct、ChatGPT 的 Dots 都能替用户订机票、订餐、下单，但越来越多的网站把它们拒之门外。最显眼的案例是亚马逊已开始封禁 Muse 的浏览与购买，而社交平台上的用户抱怨显示，类似的封禁正在各家网站上蔓延。智能体的能力在涨，网站的对抗也在涨，「替你跑腿」这门生意卡在了一个朴素的问题上：网站凭什么让机器人进门。 > 💬 上一代爬虫用 robots.txt 和网站达成了脆弱的默契，这一代智能体跳过了谈判直接进门。网站的「拒收 AI」不是技术保守，是收费站还没建好——等 agent 通行协议谈出个价格体系，墙自然会变成闸机。
来源：TechCrunch | 2026-10-06
Instinct 把智能体塞进群聊：好友无需注册，估值已到 100 亿美元
智能体创企 Instinct 宣布新功能：用户可以把它的 AI 智能体拉进群聊，帮一群人协同规划旅行、抢票、拼车，甚至分工感恩节大餐——最关键的是，群里的好友完全不需要注册 Instinct。这家公司最新一轮融资估值已达 100 亿美元。隐私设计上，群智能体与用户个人账号相互隔离，个人智能体接入群聊前需要用户批准，信任可以随时解除。最大竞对 Meta 的 Muse 目前尚无群聊功能。 > 💬 消费级智能体的下半场比的是「在场感」：不在群聊里，就不在决策现场。好友无需注册这一刀，砍的是社交流量的最后一道门槛。
来源：TechCrunch | 2026-10-05