🏆 Headline
Revolut moves 2,000 engineers off Claude Code: self-hosted Chinese open-weight models cut the bill to a tenth
One of Europe's most valuable startups is quietly swapping engines. The default AI coding environment for Revolut's roughly 2,000 engineers has moved from Anthropic's Claude Code and OpenAI's Codex to GLM, an open-weight model from Chinese lab Z.ai, running on the company's own infrastructure — first GLM-5.1, then upgraded to GLM-5.3, served by 56 Nvidia B200 GPUs. Per The Register, the company cut its AI coding inference costs by roughly 90% in under a year; David Abitbol, Revolut's AI operations and infrastructure lead, laid out the math at the AI Engineer conference in New York in October: a blended cost of about 30 cents per million tokens, roughly a tenth of what the commercial APIs charged, while rolling out AI coding tools cut coding time by about 60% and task lead time by about 70%. The road there was stepped: Gemini rolled out company-wide in early 2025 (a 20% coding-time gain, 10% on lead time), Claude Code followed in early 2026 under strict per-engineer budget caps — at the time some engineers were burning $1,000 to $10,000 a month without commensurate returns. A company with 80 million customers, a valuation around $115 billion and IPO ambitions of up to $200 billion has turned "inference is expensive" from a complaint into an engineering problem.
Source: The Register | 2026-11-03
Google's multiple-choice question for schools: turn off AI, and you lose Search too
Google announced that starting next year, AI tools for K-12 students will be off by default unless school administrators explicitly switch them on — covering Gemini, NotebookLM, Gemini in Classroom, and Google Search itself. Because AI is now woven deep into the search engine, schools that opt out of AI entirely will lose Google Search as well; students will simply see a message that the feature is unavailable at their school. Chris Phillips, Google's VP of education, put it plainly to The New York Times: "Search and AI are deeply integrated." Google says the change followed consultations with schools, teachers, and researchers. The thing you want to switch off and the thing you cannot live without now share the same pipe — this multiple-choice question has no middle option.
Source: TechCrunch | 2026-11-03
Anthropic's FlashProve: a cryptographic "signature" for model outputs, so impostors stand no chance
Anthropic published a research post on FlashProve, an open-source system in development that delivers "provable inference" — essentially a mathematical guarantee, via cryptographic proof, that a given output really came from a specific model, the way digital signatures prove an email came from its sender. The team called out two impostor scenarios: quietly routing your request to a cheaper model to cut costs, or inflating the bill by passing off a pricier one — even when the impostor is the very lab that built and operates the model.
Source: TechCrunch | 2026-11-03
Fal raises $1.3B in one round: inference demand doubles its valuation to $16B in six months
Fal, the startup that hosts and serves AI video, image, and audio models, raised $1.3 billion led by Bessemer Venture Partners with participation from Nvidia, Sound Ventures, and T. Rowe Price; per Bloomberg, the deal values Fal at $16 billion, exactly double its April valuation of $8 billion. The company announced its Series E and F at the same time, a combined $1.65 billion (Sequoia led the earlier round) — inference demand is growing so fast the rounds could not even wait to be announced separately.
Sources: TechCrunch, Bloomberg | 2026-11-03
Google cuts the firmware backdoor on 1.3 million TV boxes: the certified vendor itself was the "bandwidth broker"
Google's Threat Intelligence Group, with industry partners, dismantled a botnet known as XCPE: more than 1.3 million Android TV set-top boxes shipped with a backdoor built into their factory firmware, and users' bandwidth was quietly resold to gray-market buyers as proxies, along with fraudulent ad traffic. The most uncomfortable detail: every one of those boxes carried Google's official certification, and the backdoor came from certified vendor Videostrong — users bought the brand, and the vendor sold the users. Google publicly announced the takedown on Monday.
Sources: Ars Technica, Krebs on Security | 2026-11-03
Itch.io deletes "terabytes" of AI slop games: uploads doubled, downloads stood still
Indie game platform Itch.io is mass-purging AI-generated slop, with deletions measured in terabytes. Data published by the platform shows daily new game pages roughly doubled from just over 3,000 in early August to above 6,000 by the end of October — while downloads and purchases did not rise in step. When someone asked on Bluesky whether the platform knew why, the official account replied: "Take a guess." The pressure is industry-wide: Steam's Next Fest saw more demos than ever, with generative AI a significant factor.
Source: Kotaku | 2026-11-03
Apple draws an 18+ line for dating apps and AI companions: effective November 16
Apple told developers on Monday that apps offering dating or romantic-relationship features, and AI companions that engage in sexual or romantic content, must all carry an 18+ rating; from November 16 developers must add new content descriptors to match. It is another step after Apple's global rollout of age-assurance technology and expanded parental controls — the AI-companion category is being shelved as adult content for the first time, moving from the toy aisle to the adults-only shelf.
Source: TechCrunch | 2026-11-03
Netflix launches an AI Discover tab: find films by "mood," from scrolling lists to just asking
Netflix is rolling out a new AI Discover tab: users pick from curated themes like "something light," "scare me," "family movie night," or "quick watches," add how much time they have and what they like, and a new intelligence layer recommends results with actual clips. The streaming giant's intent is clear: replace the endless scroll with "say one sentence about your mood" and land on the target directly.
Source: TechCrunch | 2026-11-03
🏆 今日头条
Revolut 让 2000 名工程师弃用 Claude Code：自托管中国开源模型，账单砍到十分之一
欧洲估值最高的初创公司之一正在悄悄换引擎。Revolut 约 2000 名工程师的默认 AI 编码环境，已从 Anthropic 的 Claude Code 和 OpenAI 的 Codex，整体切换到跑在自家基础设施上的开源权重模型 GLM（来自中国公司智谱 Z.ai）——先上 GLM-5.1，后升级 GLM-5.3，由 56 块英伟达 B200 支撑。据 The Register 报道，不到一年，其 AI 编码推理成本下降约 90%；AI 运营与基础设施负责人 David Abitbol 在 10 月纽约 AI Engineer 大会上算的账是：混合成本约每百万 token 30 美分，约为此前商业 API 的十分之一，而铺开 AI 编码工具后编码时间缩短约六成、任务从接手到合并的交付周期缩短约七成。这条路不是一步走通的：2025 年初全公司先铺 Gemini（编码时间缩两成、交付周期缩一成），2026 年初上 Claude Code 时按人头设严格限额——当时有工程师一个月烧掉 1000 甚至 10000 美元，回报却对不上。这家拥有 8000 万用户、估值约 1150 亿美元、IPO 目标最高 2000 亿的公司，把「推理好贵」从一句抱怨变成了一道工程题。 > 💬 省钱的终点不是「换更便宜的」，是「换成自己的」。当一家按银行标准合规的公司证明了自托管开源权重够用，「默认把负载交给前沿实验室」就再也不是默认选项——实验室们真正该担心的不是降价，是客户整个离场。
来源：The Register | 2026-11-03
Google 给学校出的选择题：想关掉 AI，就得连搜索一起放弃
谷歌宣布，明年起 K-12 学生的 AI 工具将默认关闭，除非校方明确打开——覆盖 Gemini、NotebookLM、Classroom 里的 Gemini，乃至 Google 搜索本身。由于 AI 已深度织进搜索引擎，选择全面禁用 AI 的学校，学生将同时失去 Google 搜索，只会看到一条学校设置的「不可用」提示。谷歌教育副总裁 Chris Phillips 对《纽约时报》的解释很直白：「搜索和 AI 已经深度整合。」谷歌称这一调整经过了与学校、教师和研究者的磋商。想撤销的东西和离不开的东西长在同一根管道里，这道选择题没有中间项。 > 💬 捆绑从来不只是打包销售：当功能长在一条管道上，「不」就变成了需要付出代价的答案。
来源：TechCrunch | 2026-11-03
Anthropic FlashProve：给模型输出配一枚「签名」，冒名顶替无所遁形
Anthropic 发布研究博客，介绍正在开发的开源系统 FlashProve：用密码学方法实现「可证明推理」——本质上是一条数学保证，证明某段输出确实出自某个特定模型，原理类似证明邮件真的来自署名者的数字签名。研究团队特别点名了两种冒名场景：把请求偷偷转给更便宜的模型以压缩成本，或反过来用更贵的模型充数抬高身价——哪怕动手的正是运营这个模型的实验室自己。 > 💬 客户开始逐项对账的时代，「你用的到底是哪个模型」从信任问题变成了数学问题。
来源：TechCrunch | 2026-11-03
Fal 一轮吃下 13 亿美元：推理需求半年把估值翻到 160 亿
托管 AI 视频、图像、音频模型的推理服务公司 Fal 完成 13 亿美元融资，由 Bessemer 领投，英伟达、Sound Ventures、T. Rowe Price 跟投；据彭博社报道，估值达 160 亿美元，是 4 月 80 亿美元估值的整整两倍。公司还同批官宣了 E、F 两轮合计 16.5 亿美元（E 轮由红杉领投）——推理需求涨得连融资都不愿分开宣布。 > 💬 模型训练的故事讲完上半场，「把模型跑起来」正在成为下半场的主赛道。
来源：TechCrunch、Bloomberg | 2026-11-03
Google 切断 130 万台电视盒的固件后门：认证厂商自己当了「带宽贩子」
谷歌威胁情报组联合行业伙伴拆除了名为 XCPE 的僵尸网络：超过 130 万台安卓电视盒的出厂固件里被装了后门，用户的网络带宽被悄悄转卖给灰色买家当代理，还顺带刷广告流量。最扎心的细节是，这些设备全部带着谷歌官方认证，而后门出自认证厂商 Videostrong 之手——用户买的是牌子，厂商卖的是用户。谷歌已于周一公开通报切断完成。 > 💬 认证体系管得住芯片，管不住签字画押的那只手。
来源：Ars Technica、Krebs on Security | 2026-11-03
Itch.io 删掉「数 TB」AI 灌水游戏：上传量翻倍，下载量原地踏步
独立游戏平台 Itch.io 正在批量清理 AI 生成的灌水内容，删除总量以 TB 计。平台公布的数据显示，8 月初到 10 月底，每日新建游戏页从约 3000 涨到 6000 以上、整整翻倍，但下载和购买量并没有同步上涨；有网友在 Bluesky 上问平台知不知道原因，官方账号只回了一句「你猜」。同类压力正在全行业蔓延：Steam 新品节的试玩版数量创下新高，AI 生成内容是重要推手。 > 💬 生成成本趋近于零之后，「上架」不再等于「作品」，平台的筛选成本反而成了新瓶颈。
来源：Kotaku | 2026-11-03
Apple 给约会 App 和 AI 伴侣划 18+ 线：11 月 16 日起生效
苹果周一向开发者宣布 App Store 新规：提供约会或恋爱互动功能的 App，以及涉及性或恋爱情节的 AI 伴侣应用，一律要求 18+ 分级；11 月 16 日起开发者必须补齐新的内容描述标签。这是苹果在全球铺开年龄核验技术、加强家长控制之后的又一步——AI 伴侣第一次被当成成人内容来分级，这个类目正在从玩具清单挪进成人柜台。 > 💬 当一个类目需要年龄门槛，说明它已经长出了真实的情感分量——规则总比现实晚一步。
来源：TechCrunch | 2026-11-03
Netflix 上线 AI Discover 标签页：按「心情」找片，从刷列表到聊着找
Netflix 开始推出新的 AI Discover 标签页：用户从「轻松一点」「吓我一下」「家庭观影夜」「快速看完」这类策展主题中选一个，再配上可看的时长和偏好，由新的智能层推荐带正片片段的结果。这家流媒体巨头的意图很清楚：把「无限下滑刷片」换成「说一句心情」直达目标。 > 💬 推荐系统的竞争从「猜你喜欢」进入「听你说完」——对话正在变成新的搜索框。
来源：TechCrunch | 2026-11-03