🏆 Headline
OpenAI Announces Three Scientific Results: Model Disproves the Exponential Time Hypothesis, Nobel Laureate to Oversee Verification
OpenAI published a batch of major research announcements on Monday: its internal models produced one result each in mathematics, computer science, and physics — the most striking being that the Exponential Time Hypothesis (ETH) has been disproved. The conjecture, born in 2001, has been widely believed to be true, and hundreds of papers rely on it to derive conditional "this problem can't be solved fast" conclusions. OpenAI says its model found a disproof on October 16, verified in Lean the next day, with raw data and the machine-checkable proof already on GitHub. The verification setup carries even more weight: OpenAI invited the three people who proposed ETH — Impagliazzo, Paturi, and MIT's Ryan Williams — to lead an independent examination. The three promptly published an open letter drawing a line: their involvement is not an endorsement, and a result of this significance requires careful, independent scrutiny. They signed no NDAs and take no pay, and they in turn asked OpenAI to release all raw data and fund a dedicated conference — OpenAI agreed to both. The math result is equally striking: the model produced a proof of the local smoothing conjecture in every dimension, which also resolves the Kakeya conjecture and Stein's restriction conjecture that follow from it. On the physics side comes a rigorous existence proof for interacting quantum field theories in four dimensions, including QCD — with the existence of pure Yang–Mills theory being part of a Clay Millennium Prize problem — and the verification team is led by David Gross, winner of the 2004 Nobel Prize in Physics. Compare this with the episode a week ago, when OpenAI dumped over a hundred math results at once and was accused by mathematicians of "doing math by tweet." This time it handed external scholars the right to release, explain, and even decide not to publish yet — three results, three different release strategies. The discovery belongs to the models; turning it into science still belongs to people.
Source: OpenAI blog, The Next Web | 2026-10-26
Meta Plans a Raise of Up to $95 Billion: Nvidia and Friends Line Up to Buy Their Customer's Stock
Meta filed paperwork on Monday launching roughly $35 billion in two offerings — $22.5 billion of Class A common stock plus $12.5 billion of mandatory convertible preferred — with room for later issuance that takes the total to as much as $95 billion, all aimed at AI infrastructure and superintelligence. Bloomberg calls it one of the largest equity offerings in history. The most theatrical detail: Nvidia, AMD, and Broadcom have expressed non-binding interest in buying up to $5 billion of the Class A shares on the same terms as everyone else. Chipmakers subscribing to their biggest customer's stock — the money loops back into GPU orders. The market has long eyed this "circular financing" chain; now it's written into the offering documents in black and white.
Source: CNBC, Bloomberg, Business Insider | 2026-10-26
Anthropic Officially Files Its S-1: $16.3 Billion First-Half Revenue, an $80.7 Billion Net Loss
Anthropic formally filed for its IPO on Monday under the ticker "ANT." The prospectus shows: $16.3 billion of revenue in the first half of 2026, more than thirteen times the $1.2 billion a year earlier; operating losses of $693 million, clearly narrowed from $2.1 billion; and a net loss of $80.7 billion — mostly one-time, non-cash accounting adjustments tied to convertible debt. The enterprise base is solid: more than 11,900 customers spending over $100,000 a year, 1,900 above $1 million, and roughly 250 above $10 million. The governance setup stays on brand: the seven co-founders control 50.1% of voting power through a Founder LLC, and the Long-Term Benefit Trust appoints most of the board. Offering size and pricing are undisclosed. Last week OpenAI was negotiating a $30 billion round while analysts dissected "strange pre-IPO pricing"; this week Anthropic handed in its paperwork — the IPO race is officially in its final seconds.
Source: TechCrunch, Fortune, CNBC | 2026-10-26
SpaceX Locks In $40 Billion of GPU Leases: Apollo Buys the Chips, SpaceX Rents Them Back
SpaceX disclosed in a regulatory filing on Monday a binding commitment of roughly $40 billion for AI chips and data center equipment — which the company calls the "largest committed GPU financing." The structure is a classic sale-leaseback: an Apollo-managed entity buys the chips and servers and leases them to SpaceX, on leases of about five years, closing by year-end with equipment arriving through the first half of 2027. Musk warmed up the crowd on social media over the weekend, noting the company carries an investment-grade rating with revenue from Starlink and launch. The backdrop is more striking: the AI business — which the company calls "Super Intelligence" — posted $7.56 billion in quarterly revenue, up ninefold year over year, now its largest segment. Musk also predicted that of two hundred-plus "neocloud" providers, most face widespread failures and consolidation.
Source: TechCrunch, CNBC | 2026-10-26
GPT-6.1 Terra Launches: Built for Computer Use, Half the Price, About a Tenth the Cost per Task of the Flagship
OpenAI released GPT-6.1 Terra, positioned as "fast and efficient": visual reasoning, precise computer use, and more natural writing, aimed at the workday — filling forms, entering invoices, editing videos. Standard API pricing is half the previous generation: $1 per million input tokens and $5 per million output tokens. The scorecard: on a benchmark of professional tasks, medium-reasoning beats the previous generation by 15 points at about a third the cost, matches flagship GPT-6 Astra at a tenth of the price, and finishes tasks over three times faster than Astra and Sol. Hallucination control is another highlight: the low-effort error rate fell from 53.6% to 16.7%. Available immediately to full-subscription users. After last week's pre-IPO pricing timeline analysis, OpenAI keeps pushing "cheap and good enough to work" onto enterprise purchase orders — with real prices this time.
Source: OpenAI blog, The Next Web | 2026-10-26
Cloudflare Discloses a KV-Cache Cross-Read Incident: 155 Long-Context Responses Carried Other Users' Inputs
Cloudflare published an incident report: on its Workers AI platform, the GLM-5.3 model served a small number of users responses containing content from other users' requests "made shortly before theirs" between late August and early September, due to a flaw in the caching mechanics of its inference engine. The scope: 155 responses, all from prompts longer than roughly 680,000 tokens. Cloudflare says that after excluding its own tests and the researcher's requests, the remaining traffic's timing and volume are consistent with ordinary long-context use, with no evidence of deliberate exploitation; the affected model was remediated in early September, no other models were impacted, and the company coordinated disclosure with SGLang, the open-source inference engine involved, hardening production deployments. The cheaper long context gets and the more aggressive shared caching becomes, the more every API user should weigh the risk of their prompts leaking through someone else's session.
Source: Cloudflare blog | 2026-10-26
An Instruction File in Your Downloads Folder Can Hijack Coding Agents: 90% Success in Tests
Security firm Forever Security disclosed that coding agents automatically read instruction files in their working directory at startup — and miscreants plant doctored AGENTS.md files in the Downloads folder, waiting for you to run an agent there so it follows the fake instructions. The firm caught exactly such a file on a real customer's machine, instructing the agent to install a package impersonating legitimate software; in reproduction tests, launching Codex from Downloads executed the test payload in roughly 90% of trials. The Claude Code demonstration was more direct: a forged instruction file could harvest login credentials and take over the account. The researchers' advice is plain: don't use Downloads as an agent's working directory, and check the folder for instructions of unknown origin before running tasks.
Source: BleepingComputer | 2026-10-25
Nvidia Mulls a Software Fee: About $2,000 per GPU per Year to Run Agents
Jensen Huang offered a new number at GTC Berlin: licensing Nvidia's software to run agents in "AI factories" costs about $2,000 per GPU per year. A third-party extrapolation: at five million subscribed GPUs, that's $10 billion a year, $50 billion over five years. The wrinkle: Nvidia's existing enterprise software platform lists at $4,500 per GPU per year — whether the new figure is a discount or a different scope, the company hasn't said, nor when it would take effect. Huang reiterated the transformation ambition: "One of these days, people might not categorize us as a chip company." Two days earlier, Nvidia shipped a reference architecture that builds security fences into enterprise infrastructure, with Anthropic and Salesforce already announcing integrations.
Source: Tom's Hardware | 2026-10-25
🏆 今日头条
OpenAI 宣布三项科学成果：模型证伪指数时间假设，诺奖得主带队核验
OpenAI 周一发布一批重磅科研公告：内部模型在数学、计算机科学、物理学三个领域各产出一项成果，其中最劲爆的是——指数时间假设（ETH）被证伪了。这条假设诞生于 2001 年，被业界广泛认为是成立的，数百篇论文靠它推导「某问题不可能算得快」的条件性结论；如今 OpenAI 说，自家模型 10 月 16 日找到了反证，次日通过 Lean 形式化验证，原始数据和机器可验证证明已经放上 GitHub。更有分量的是核验安排：OpenAI 请来了 ETH 的三位提出者——Impagliazzo、Paturi 和麻省理工的 Ryan Williams——牵头独立审查。三人随即发公开信划清界限：参与不构成对结论的背书，这种量级的结果需要审慎独立的检验；他们不签保密协议、不拿报酬，还反过来要求 OpenAI 公开全部原始数据、出资一场专门会议，OpenAI 全都答应了。数学侧的成果同样惊人：模型给出了局部光滑猜想的全维度证明，顺带解决了由它推出的 Kakeya 猜想与 Stein 限制猜想；物理侧则是一个四维量子场论（含 QCD）的严格存在性证明，其中纯杨–米尔斯理论存在性正是千禧年大奖难题的一部分，核验团队由 2004 年诺贝尔物理学奖得主 David Gross 领衔。对比一周前的风波——当时 OpenAI 一次性甩出上百个数学结果，被数学界指责「发推文做数学」——这一次它把发布权、解释权、甚至「先不公开」的决定都交给了外部学者，三种成果用了三种不同的发布策略。发现是模型的，消化成科学，还得靠人。 > 💬 上次是甩结果，这次是交流程。证伪 ETH 若坐实，几百篇论文的「不可能」要重写；但比结论更值得盯的是这套「牧羊人」机制——发现归模型，背书归人类，而且背书的人有权说不。AI 产出越接近科学前沿，「谁来验」就会越值钱。
来源：OpenAI 官方博客、The Next Web | 2026-10-26
Meta 拟最高募资 950 亿美元：英伟达们排队认购自家客户的股票
Meta 周一提交文件，宣布启动约 350 亿美元的两笔发行——225 亿美元 A 类普通股加 125 亿美元强制可转换优先股——并预留后续发行通道，总融资规模最高达 950 亿美元，全部投向 AI 基础设施与超级智能。彭博社称这是史上最大规模的股权发行之一。最有戏剧性的一笔是：英伟达、AMD、博通三家芯片厂已表达非约束性意向，合计认购至多 50 亿美元 A 类股，条件与其他投资者完全相同。芯片厂掏钱认购大客户，钱转一圈又变回 GPU 订单——这条「循环融资」的链条，市场已经盯了很久，如今被白纸黑字写进了招股文件。 > 💬 供应商认购客户的股票，等于自己给自己的销量做担保。产业链绑成这样，涨的时候是合力，回头看的时候，怕是要算谁在给谁抬轿子。
来源：CNBC、Bloomberg、Business Insider | 2026-10-26
Anthropic 正式递交 S-1：上半年营收 163 亿美元，净亏损 807 亿
Anthropic 周一正式递交 IPO 申请，代码「ANT」。招股书披露：2026 上半年营收 163 亿美元，是去年同期 12 亿的十三倍多；经营亏损 6.93 亿美元，较上年同期 21 亿明显收窄；但净亏损高达 807 亿美元——主要是可转债相关的一次性非现金会计调整。企业基本盘相当扎实：年付超 10 万美元的客户超过 11,900 家，超 100 万的 1,900 家，超 1,000 万的约 250 家。治理结构延续一贯画风：七位联合创始人通过 Founder LLC 控制 50.1% 投票权，长期利益信托有权任命多数董事。发行规模与定价未披露。上周是 OpenAI 谈 300 亿美元新融资、分析师拆解「上市前奇怪定价」；这一周，Anthropic 把文件递了上去——两家的IPO竞赛正式进入读秒。 > 💬 十三倍营收增速配上 807 亿账面亏损，这份招股书就是给华尔街出的一道定价题：你按成长股算，还是按亏损算？创始人们早有答案——投票权一分不让，钱照拿，风险交给公开市场。
来源：TechCrunch、Fortune、CNBC | 2026-10-26
SpaceX 锁定 400 亿美元 GPU 租约：Apollo 买卡，星舰还租
SpaceX 周一在监管文件中披露，已与 Apollo 管理的实体达成约 400 亿美元的约束性承诺，用于购置 AI 芯片与数据中心设备——公司称之为「最大规模的 GPU 融资承诺」。结构是标准的「购后回租」：Apollo 买下芯片和服务器，再租给 SpaceX，租期约五年，年底前交割，设备 2027 年上半年陆续到位。马斯克周末先在社交平台预热，称公司已获投资级评级，收入来自星链和发射业务。数字背景更惊人：AI 业务（公司称「超级智能」）上季度营收 75.6 亿美元，同比涨九倍，已成公司最大业务板块；马斯克还预言，两百多家「新云」厂商将迎来大规模出清和整合。 > 💬 一家火箭公司成了 GPU 资本运作最激进的主角。私募信贷进场接住了芯片账单，租约把一次性天价采购摊成五年租金——AI 军备竞赛的融资工程学，越来越像当年的铁路和航空。
来源：TechCrunch、CNBC | 2026-10-26
GPT-6.1 Terra 发布：主打「电脑操作」，价格砍半，单任务成本约为旗舰十分之一
OpenAI 深夜发布新模型 GPT-6.1 Terra，定位是「快速高效」：视觉推理、精确的电脑操作加上更自然的写作，瞄准填表、录入发票、改视频这类日常工作。标准 API 价格比上代再降一半——输入 1 美元、输出 5 美元每百万 token。账面成绩单：在专业任务基准上，中档推理比上代高 15 个百分点、成本约三分之一；与旗舰 GPT-6 Astra 打平、单价只有十分之一；任务完成速度比 Astra 和 Sol 快三倍余。幻觉控制是另一大亮点：低推理档错误率从 53.6% 直落到 16.7%。对全套订阅用户即刻可用。继上周的降价时间线分析之后，OpenAI 用实价继续把「便宜好用的干活模型」往企业采购单上推。 > 💬 不追最聪明，专攻「够用且便宜」——单任务成本压到旗舰十分之一，这是冲着「让每张办公桌都养一个数字员工」去的。旗舰秀肌肉，Terra 走量，定价表比发布会诚实。
来源：OpenAI 官方博客、The Next Web | 2026-10-26
Cloudflare 披露推理缓存串读事故：155 条长文响应把别人的输入带给了别人
Cloudflare 发表事故报告：其 Workers AI 平台上的 GLM-5.3 模型，8 月底到 9 月初因推理引擎的缓存机制缺陷，把部分用户收到的响应里混入了其他用户「在不久前发出的请求」相关内容。影响范围：155 条响应，全部来自长度超过 68 万 token 的超长提示。Cloudflare 称已排除自身测试与发现者的请求，其余流量的时序和体量均与正常长文使用吻合，未发现有人蓄意利用；涉事模型 9 月初已处置，其余模型不受影响，并与所用的开源推理引擎 SGLang 做了协调披露，生产部署已加固。长上下文越便宜、共享缓存越激进，这类「你的提示词从别人的会话里漏出来」的风险就越值得每个 API 用户掂量。 > 💬 为了省算力，把不同用户的缓存放在同一口锅里——省下的钱和漏出去的隐私，这笔账总算被摆上了台面。共享缓存的性能红利是真的，隔离边界也一样是真的。
来源：Cloudflare 官方博客 | 2026-10-26
「下载」文件夹里的指令文件能劫持编码智能体：九成试验成功
安全公司 Forever Security 披露：编码智能体启动时会自动读取工作目录下的 AGENTS.md 指令文件——而不法分子把做了手脚的这类文件放进「下载」文件夹，等你让智能体在这里干活，它就会照着假指令办事。该公司在真实客户机器上就抓到一份这样的文件，指示智能体安装冒充正常软件的包；复现测试里，从「下载」文件夹启动 Codex，九成试验执行了测试载荷。对 Claude Code 的演示更直接：伪造的指令文件能骗取登录凭据、接管账号。研究者的建议很朴素：别把「下载」文件夹当智能体的工作目录，跑任务前先看一眼目录里有没有来路不明的指令文件。 > 💬 智能体的世界观是「读到的都信」，而「下载」文件夹恰是互联网上最脏的抽屉。给智能体划工作目录，就是给它划教养的边界——这条建议值一次账号。
来源：BleepingComputer | 2026-10-25
英伟达盘算软件年费：每张 GPU 每年 2,000 美元，跑智能体的软件全包
黄仁勋在 GTC 柏林站上给出一个新说法：用英伟达软件在「AI 工厂」里运行智能体，许可费大约是每张 GPU 每年 2,000 美元。第三方推演：若 500 万张 GPU 订阅，一年就是 100 亿美元，五年 500 亿。微妙的是，英伟达现有的企业软件平台挂牌价是每张 GPU 每年 4,500 美元——新数字是更低还是覆盖范围不同，公司没有说清，生效时间也没定。黄仁勋同时重申转型的野心：「有一天，人们可能不再把我们归类为芯片公司。」就在两天前，NVIDIA 刚发布把安全围栏做进企业基础设施的参考架构，Anthropic 和 Salesforce 都已宣布接入。 > 💬 卖卡变成收租，一次买断变成终身订阅——这是所有硬件公司的终极梦想，也是所有客户的终极噩梦。每张卡 2,000 美元的「智能体税」真落地，下一个被讨论的就是「离开英伟达」的成本了。
来源：Tom's Hardware | 2026-10-25