🏆 Headline
Google announces Gemini 4 Argon: 1M-token output, but you can't use it yet
Google announced its new flagship model, Gemini 4 Argon, late on September 30 — officially positioned as a frontier model fighting on three fronts: real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense. The hard numbers: output token limit jumps from 64K to an industry-leading 1M tokens; input pricing is $2 per million tokens and output $10, with cached input at 95% off. It scores 77.9% on the DeepSWE v1.1 software-engineering benchmark, which per Google's own table places it above GPT-6 Astra, Fable 5.1 and Opus 5.5. The internal war stories are even flashier: Argon agents scanned fleet-wide profiling telemetry across Google's data centers and autonomously applied memory optimizations, freeing over 300 TiB once rolled out (estimated total savings of 500 TiB to 1 PiB). Another set of agents is migrating Google's internal C/C++ codebases to Rust — from tens of thousands of lines in core libraries like re2 and libgav1 up to 800K+ lines in the Fuchsia OS Zircon kernel; for libgav1, Argon rewrote 32K lines of SIMD code so the compiler could auto-vectorize, producing a memory-safe video decoder 2.7x faster than the original Rust port. But the real center of this story is how it's being released: ordinary users can't have it yet. Google is giving Argon only to a small group of trusted cyber defenders through its Fairwind Program (Wiz is already using it — the company says the model surfaced a critical flaw in hospital software worldwide that previous frontier models had missed). API and consumer access will follow "as soon as possible," with no timeline. The stated rationale is that frontier capability demands a phased rollout, in line with the U.S. government's voluntary pre-release access process. And the same day, Bloomberg poured cold water on the party: some Google employees question how well Argon performs in key areas like coding based on internal testing, and the company is grappling with that internal skepticism; Google's side responded that there is "large consensus" internally that the model is at the frontier and denied that it struggles with real-world coding tasks. Choosing cyber defenders as the first audience anchors the narrative in safety — except this time, the safety narrative has to hold up not just against outside critics, but against Google's own employees.
Source: Google official blog, Ars Technica, Bloomberg | 2026-09-30
Reddit sentences RSS to death: feeds end November 13, public API closes in March 2027
On September 30, Reddit announced it is winding down RSS feed support, with the official line that RSS has become "a common surface for large-scale scraping and automated abuse": RSS support ceases on November 13, and the public API shuts down in March 2027, with a batch of replacement options for moderators and developers attached to the announcement. RSS was a symbol of the blogging-era "open web" — a content outlet that needed no account, no approval, and was machine-readable. When even Old Reddit's stripped-down interface is being restricted for inactive users, this twenty-year-old open channel is officially in its countdown.
Source: TechCrunch | 2026-09-30
DeepSeek and Huawei publicly release Ascend chip programming tools, aiming at Nvidia's software moat
On September 30, DeepSeek announced a partnership with Huawei and publicly released programming tools optimized for Ascend chips. Reuters framed it as another sign of Chinese tech firms deepening ties as they seek alternatives to the Nvidia ecosystem; Bloomberg's headline was blunter: tools that "may replace Nvidia's." Nvidia's true moat was never just the chips — it's two decades of software built on top of CUDA. DeepSeek, one of the best-known Nvidia customers, is now laying bricks for another software stack. Chips can be cut off in one policy move or another; the migration of a software ecosystem is the real watershed.
Source: Reuters, Bloomberg, NYT | 2026-09-30
Google's pilot paying websites for AI answers is reported as struggling: most publishers get one-thousandth of their ad revenue
Ars Technica reported on September 30, citing The Information, that Google's AI contribution pilot has admitted about 100 publishers — sites get paid when their content materially contributes to Gemini-powered answers like AI Overviews. But most small and mid-sized publishers are receiving amounts equal to roughly 0.1% of their advertising revenue, and several larger publishers have declined to join, hoping absence forces a more decent offer. The distribution is extremely lopsided: one early joiner is on track to earn more than $1 million a year, a recent addition has earned $50,000 to $60,000, and some smaller sites have seen less than $1,000 over several months. Publishers also complain settlement is opaque — nobody can say what counts as a "meaningful contribution," so earnings can't be planned. Background: NYT's copyright lawsuit against Google is ongoing, and the European Commission's antitrust probe is examining whether Google fairly compensates publishers for AI use of their content.
Source: Ars Technica, The Information | 2026-09-30
OpenAI's next round surfaces: at least $30B at roughly a $1.4T valuation
Bloomberg reported on September 29 that OpenAI is in talks with investors to raise at least $30 billion in a pre-IPO round at a valuation of roughly $1.4 trillion; investors are eager to pile in before next year's expected public debut. Per the report, run-rate revenue has jumped 70% since July, reaching $40 billion in August; the previous round was $122 billion in March at an $852 billion valuation, originally intended as the last private raise before an IPO. Separately, per Ars Technica, CEO Sam Altman has ruled out going public in 2026, prioritizing safety first — he told Fortune he couldn't accept attaching his name to an endeavor with "like a 10% chance of killing everybody by the end of the decade." A $1.4 trillion valuation negotiation and a postponed IPO being true at the same time is itself the most honest footnote of this phase.
Source: Bloomberg, TechCrunch, Ars Technica | 2026-09-29
ElevenLabs' valuation is now official: $22B, double in a year
Voice AI company ElevenLabs confirmed on September 30 that its new $300 million employee tender offer was co-led by Wellington and T. Rowe Price, valuing the company at $22 billion — twice the $11 billion valuation from its $500 million raise in February. It's the company's second employee-liquidity move in a year: last September's tender was at a $6.6 billion valuation, meaning the valuation has more than tripled in twelve months. Employee liquidity is becoming standard retention equipment for AI startups: let early employees cash out, in exchange for staying at the table. Today's news confirms the $22 billion figure reported as investor chatter on September 25.
Source: TechCrunch | 2026-09-30
Coding-startup infighting goes public: Factory's CEO accuses his own VC board advisor of spying for rival Cognition
Factory CEO Matan Grinberg announced on X on September 30 that he had removed venture investor Chris Degnan from his board-advisor role, alleging Degnan shared confidential information with Factory's biggest competitor, Cognition. Two dramatic hours later, Degnan announced he had joined Cognition as chief revenue officer, then pushed back that he wasn't fired — he resigned when he told Grinberg he was taking the job. Degnan was Snowflake's first sales hire and spent 11 years as its chief revenue officer; for the past five months he had been Factory's board advisor via RPT Partners. Competition in AI coding has escalated from fighting over customers to fighting over people with no politeness left at all.
Source: TechCrunch | 2026-09-30
DoorDash launches a text-to-order agent: say "my usual" in Apple Messages and dinner is handled
DoorDash announced on September 30 an ordering agent inside Apple Messages: send a prompt like "order my usual" and the agent understands you mean your Friday-night regulars, searches local spots, builds a cart around dietary preferences, and even texts photos of the dishes it recommends; group orders can mix dietary restrictions and quantities in one go. It opens with a U.S. waitlist. DoorDash's play is to outrun Uber Eats and Grubhub, but the more notable part is the trend: after Google embedded Flipkart checkout inside Gemini and Shopify opened a checkout lane for browser agents, another class of everyday consumption is moving out of apps and into conversation windows.
Source: TechCrunch | 2026-09-30
🏆 今日头条
Google 发布 Gemini 4 Argon：输出上限 100 万 token，但普通用户还用不上
Google 9 月 30 日深夜发布新一代旗舰模型 Gemini 4 Argon，官方定位是「真实世界软件工程、法务金融等企业知识工作与网络防御」三线作战的前沿模型。几个硬数字：输出 token 上限从上一代的 6.4 万直接拉到 100 万（行业领先），输入价格 2 美元/百万 token、输出 10 美元/百万 token，缓存输入打到 95 折扣；DeepSWE v1.1 软件工程基准拿下 77.9%，按官方表格高于 GPT-6 Astra、Fable 5.1 与 Opus 5.5。内部战绩更抢眼：Argon 智能体扫描 Google 全数据中心的分析数据，自主完成内存优化，上线后释放超过 300 TiB 内存（预计总节省最高约 1 PiB）；另一组智能体在把 Google 内部 C/C++ 代码库迁移到 Rust，规模从 re2、libgav1 等核心库的数万行一路到 Fuchsia OS Zircon 内核的 80 万行以上，其中 libgav1 的 Rust 版视频解码器经它重写后比原 Rust 移植快 2.7 倍。 但这条新闻真正的重心是发布方式：普通用户现在用不上。Google 只把 Argon 交给「Fairwind 项目」的一小批网络防御者（Wiz 已在用，官方称它在一个全球医院在用的医疗软件里发现了此前前沿模型都漏掉的高危漏洞）， API 与消费者版「尽快」跟进，没有时间表。官方给的理由是前沿能力需要分阶段发布，并配合政府的预发布自愿流程。而同一天彭博社泼了盆冷水：部分 Google 员工在内部测试中怀疑 Argon 在编码等关键场景的真实表现，称公司内部对此存在分歧；Google 一方则回应称内部有「高度共识」认为模型处于前沿、否认真实编码任务吃力。发布会选在网络防御者先行，恰好把叙事锚在安全上——只是这次，安全叙事要顶住的不只是外部质疑，还有自家员工的。 > 💬 连日风暴之后，行业的新共识是能力不再是发布的唯一变量，「先给谁用」才是。Google 这次把顺序写成了一份声明：防御者先于开发者、开发者先于消费者。历史不会忘记，上个月 OAI 也是在发布当天被曝「内部还没理清」——旗舰模型的发布正在从技术事件变成信任事件，员工的私下评价都能左右股价，这场戏才刚开场。
来源：Google 官方博客、Ars Technica、Bloomberg | 2026-09-30
Reddit 宣判 RSS 死刑：11 月 13 日停服，公共 API 明年 3 月关停
Reddit 9 月 30 日宣布逐步停止 RSS 订阅源支持，官方口径是 RSS 已成为「大规模抓取和自动化滥用的常见入口」：RSS 将于 11 月 13 日停止服务，公共 API 则定在 2027 年 3 月关停，公告同时给版主和开发者附带了一批替代方案。RSS 是博客时代「开放网络」的象征——一个无需账号、无需审批、机器可读的内容出口。当连「老 Reddit」的简化界面都要限制不活跃用户使用时，这条二十多年的开放通道正式进入倒计时。 > 💬 拦爬虫是真需求，但药方是把「不需要登录就能读」的通道全部焊死。开放网络不是被某个技术淘汰的，是被一次次「为了安全」的收口磨掉的。
来源：TechCrunch | 2026-09-30
DeepSeek 携手华为公开昇腾芯片编程工具，剑指英伟达的软件护城河
DeepSeek 9 月 30 日宣布与华为合作，公开发布为昇腾芯片优化的编程工具。路透社指出，这是中国科技公司深化合作、寻求英伟达生态替代方案的又一信号；彭博社的标题更直白：这批工具「可能替代英伟达的」对应软件。英伟达真正的护城河从来不只是芯片，而是 CUDA 之上二十年的软件栈——DeepSeek 是最知名的英伟达客户之一，如今亲手给另一套软件栈添砖。芯片可以断供来断供去，软件生态的迁移才是真正的分水岭。 > 💬 卖卡的人最怕的不是别人造出一张卡，而是别人造出一整套「不用你家卡也能干活」的工具箱。这一步的象征意义大于当下的性能数字。
来源：Reuters、Bloomberg、NYT | 2026-09-30
Google 给网站付「AI 答案引用费」试点被曝遇冷：多数出版商只拿到广告收入的千分之一
Ars Technica 9 月 30 日援引 The Information 报道，Google 的 AI 内容贡献试点已吸纳约 100 家出版商——网站内容对 AI Overview 等 Gemini 驱动的答案有实质贡献时可获付费。但多数中小出版商到手的金额约等于自身广告收入的 0.1%，几家大出版商已拒绝加入，想用缺席逼一个更体面的报价。分配也极不均：早期加入的一家年入有望超 100 万美元，最近加入的一家只拿到 5-6 万美元，更有小型站几个月累计不足 1000 美元。出版商还抱怨结算不透明——没人说得清什么算「实质贡献」，根本没法规划。背景是 NYT 对 Google 的版权诉讼仍在进行，欧洲委员会也在反垄断调查中审视 Google 对出版商的补偿是否公平。 > 💬 「内容有价」喊了一年，落地是个千分之零点一的零钱。流量分成的旧契约正在失效，新契约连草稿都还没写完——大玩家用退场议价，小玩家只能看着 AI 摘要替他们把读者截留在搜索页。
来源：Ars Technica、The Information | 2026-09-30
OpenAI 新一轮融资谈判曝光：至少 300 亿美元、估值约 1.4 万亿美元
彭博社 9 月 29 日报道，OpenAI 正与投资人洽谈一轮至少 300 亿美元的上市前融资，估值约 1.4 万亿美元；投资人急着在明年上市前再上一车。据披露，公司 7 月以来 run-rate 收入增长 70%，8 月已达 400 亿美元；上一轮是 3 月的 1220 亿美元融资、投后估值 8520 亿美元，原本计划是 IPO 前最后一轮私募。另据 Ars Technica，CEO 奥尔特曼已排除 2026 年上市的可能，理由是「优先把安全做好」——他对《财富》说过，不能接受「十年内有 10% 概率杀死所有人」的事业挂着自己的名字。一边是 1.4 万亿美元的估值谈判，一边是推迟上市的表态，两件事同时为真，本身就是这个阶段最真实的注脚。 > 💬 估值先于上市到达终点线，IPO 变成了流程而非里程碑。当一级市场的钱已经多到不董事程，检查的就不是公司值多少钱，而是这场排队里还有多少新钱。
来源：Bloomberg、TechCrunch、Ars Technica | 2026-09-29
ElevenLabs 估值官宣落定：220 亿美元，一年翻倍
语音 AI 公司 ElevenLabs 9 月 30 日确认，新一轮 3 亿美元的员工股回购（tender offer）由 Wellington 与 T. Rowe Price 联合领投，公司估值按 220 亿美元计算——是 2 月上一轮 5 亿美元融资（110 亿美元估值）的两倍。这是它一年内第二次员工流动性操作：去年 9 月的回购按 66 亿美元估值，也就是说一年估值翻了三倍多。员工流动性正在成为 AI 创企的标配留人工具：让早期员工先落袋，换取他们继续留在牌桌上。昨天的更新对应 9 月 25 日报道中「投资人给出 220 亿美元估值」的传闻，如今数字坐实。 > 💬 一年翻三倍的速度里，装着语音这个赛道从「配套功能」变成「入口级产品」的全程。留人靠期权不靠情怀，这场人才战一级市场替二级市场先打了。
来源：TechCrunch | 2026-09-30
编程创业公司内讧上桌：Factory CEO 公开指控自家董事顾问为竞争对手 Cognition 当间谍
Factory CEO Matan Grinberg 9 月 30 日在 X 上宣布，已解除风险投资人 Chris Degnan 的董事顾问职务，并指控 Degnan 向 Factory 最大的竞争对手 Cognition 泄露机密信息。戏剧性的两小时后，Degnan 宣布加入 Cognition 担任首席营收官，随后反驳称自己不是被炒、而是告知 Grinberg 自己要去 Cognition 时主动辞职。Degnan 是 Snowflake 首位销售员工、当了 11 年首席营收官，近五个月以 RPT Partners 合伙人身份担任 Factory 的董事顾问。AI 编程这条赛道的竞争烈度，已经从抢客户卷到了抢人时连体面都不留。 > 💬 硅谷的董事会政治第一次以「抓内鬼」的形式公开直播。编码智能体的钱有多热，桌子底下的动作就有多密——这大概是行业过热的另一个佐证。
来源：TechCrunch | 2026-09-30
DoorDash 上线「发短信点外卖」智能体：Apple Messages 里说一句「老规矩」就能下单
DoorDash 9 月 30 日宣布在 Apple Messages 里上线点餐智能体：用户发「order my usual」这样的提示，智能体就能理解你指的是周五晚常点的那几样，还会搜本地餐厅、按口味拼购物车，甚至把推荐菜品照片直接发进对话；多人订单能一次处理不同的忌口与数量。功能先开美区等候名单。DoorDash 的算盘是对 Uber Eats 和 Grubhub 抢跑，但更值得注意的是趋势：继 Google 在 Gemini 里内嵌 Flipkart 购物、Shopify 给浏览器智能体开结账通道之后，又一类日常消费动作从 App 里搬进了对话窗口。 > 💬 「老规矩」三个字是这类产品的真正技术难点——它要的不是听懂菜名，而是记住你。当平台开始争抢对话入口，App 时代的交互地基就松了一块。
来源：TechCrunch | 2026-09-30