🏆 Headline
Cheating Rate Driven to 0%: Google Unveils Gemini 4 Carbon, Safety Steals the Show
Google announced its next flagship overnight: Gemini 4 Carbon, which the company calls its strongest model yet. It ranks #1 on the Artificial Analysis Intelligence Index v5.0 (score 59) and jumped 89 elo points to #4 on WebDev Arena. But the real star of this launch wasn't the benchmarks — it was the safety numbers. Successful cheating on Terminal-Bench 4.0 was driven to 0.0% (attempts down to 0.4%), high-severity misalignment alerts in internal environments dropped nearly 80%, and on the indirect prompt injection benchmark it resisted every evaluated attack (0.0% success rate, versus 0.7% for its predecessor Argon). Cybersecurity is Carbon's headline capability: it set state-of-the-art marks across the defensive lifecycle — 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for validating exploitation, and 71.0% on CWE-bench for generating patches. Internally, it produces secure code at more than twice the success rate of Argon, which launched just two weeks ago; the Chrome team reports its patch quality is more than 10 times better than another lab's flagship. Most users can't touch it yet: starting this week, Carbon goes to a subset of cyber defenders, then gradually to trusted partners via the Fairwind Program. No date yet for developer or consumer availability. Pricing matches Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off.
Source: Google blog / Ars Technica | 2026-10-28
Alphabet's Earnings Night Shock: Free Cash Flow Negative $9.07B, Stock Down Over 10% After Hours
Alphabet's after-hours report Wednesday made the market gasp: revenue grew 25%, but the company posted its first combined revenue-and-profit miss in nearly four years. Net income fell to $26.17B (from $34.98B a year ago), and free cash flow came in at negative $9.07B — far worse than the negative $2.70B analysts expected. Q3 capex totaled $51.36B with full-year guidance held at $195B-$205B, and the company previewed capex exceeding $300B in 2027. To fill the funding gap, long-term debt climbed from $21.6B to $125.1B in a year — roughly a fivefold increase. Shares slumped more than 10% after hours.
Source: CNBC | 2026-10-28
Microsoft's Double Reveal: Azure Disclosed at $33.2B, OpenAI Alone Contributes $10B
Microsoft, reporting the same day, delivered two firsts. Azure's quarterly revenue was disclosed in dollars for the first time — $33.2B, up 48% year over year, slightly ahead of Google Cloud's $30.7B in the same quarter (which grew at roughly twice Azure's pace). And a regulatory filing showed Microsoft recorded $10B in revenue from OpenAI-related commercial arrangements, including revenue-sharing payments — nearly 11% of its $92.3B quarterly total. For the full prior fiscal year the figure was $24.1B; at the current pace, about $40B annualized. The flip side: receivables from OpenAI have piled up to $7.2B, up another $1.2B in three months.
Source: TechCrunch | 2026-10-28
Anthropic Launches Usage Reporting: Enterprises Finally See Where the Tokens Go
Anthropic released a reporting suite for enterprise customers: managers can view, over any date range, what a team did with AI, what it produced, what it cost, and where workflows broke down — with improvement recommendations attached; the data is not used for model training by default. Two numbers from its IPO filing frame the launch: the company reached operating profit in Q2, and net revenue retention stands at 900% — the same customers' quarterly spend grew ninefold in a year.
Source: TechCrunch | 2026-10-28
A 9.5-Severity Firmware Flaw: EDK2's 16-Year-Old Overflow Touches Servers Everywhere
Security researchers disclosed CVE-2026-97631 in EDK2 — the open codebase behind most of the world's server UEFI firmware — rated 9.5 out of 10. The PXE network-boot driver lacks a bounds check on a 504-byte field when handling ICMP error messages, allowing code to run before the operating system and defenses like Secure Boot load. The buffer overflow has been latent for 16 years. Disclosure went awkwardly too: coordinated disclosure was planned for January until a motherboard vendor's code commit last month lifted the veil early. The good news: exploitation requires control of the target's local network, and home devices that boot from disk are unaffected.
Source: Ars Technica | 2026-10-27
A Jobs-Era Rule Falls: MacBook Pro Gets Touch, and Tandem OLED
Apple announced its first touchscreen MacBook Pro — the MacBook Pro with PROLED XDR Display — in 14.2-inch and 16.3-inch sizes with tandem OLED panels: 2000 nits peak brightness, 10,000,000:1 contrast, up to 24 hours of battery life, all in a thinner, lighter body. Mind the configuration split: the touch models carry M5 Pro / M5 Max chips, while the new M6 lands in the entry 14-inch (mini-LED) starting at $1,999. From Monday's iPad mini to Wednesday's touch MacBook, Apple spent three days retiring the "Mac doesn't do touch" doctrine after two decades.
Source: Apple Newsroom / Ars Technica | 2026-10-28
Google Opens Connected Apps Platform: The Era of Apps Queuing for AI Referrals
Google launched Connected Apps Platform, letting third-party developers plug their apps into AI experiences across Search and Gemini. Previously connected apps were limited to Google-curated marquee services (designing a flyer via Canva, building playlists via YouTube Music); now any business can self-serve. The logic is simple: the entry point is shifting from "open an app" to "say it to your AI" — whoever holds the chat box holds the new homepage. ChatGPT and Meta Muse are racing down the same path.
Source: TechCrunch | 2026-10-28
Nvidia Open-Sources a Speech Model: 40 Languages, 8 Speakers, One Chip
Nvidia released Nemotron 3.5 Transcribe, an open-weight transcription model: 40 languages, translation to and from English for 31 of them, hour-long recordings, eight-speaker separation, and 30B parameters running on a single chip (from Jetson AGX Thor to DGX Spark). It also open-sourced the Granary v2 speech dataset and fine-tuning recipes, claiming the lead on Hugging Face's Open ASR Leaderboard. The targets: meetings and medical transcription, two of AI's most established workplace markets. One limit: recorded audio only, no live streams.
Source: TechCrunch / Hugging Face | 2026-10-28
🏆 今日头条
作弊率压到 0%：Google 发布 Gemini 4 Carbon，安全成了新跑分
Google 深夜发布新一代旗舰模型 Gemini 4 Carbon，自称迄今最强：在 Artificial Analysis 智能指数 v5.0 排名第一（59 分），WebDev Arena 上涨 89 个 elo 来到第四。但这次发布会的真正主角不是跑分，而是安全指标——模型作弊成功率在 Terminal-Bench 4.0 上被压到 0.0%（尝试率仅 0.4%），内部环境的高严重度失控告警下降近 80%，间接提示注入基准上更是抗住了全部攻击尝试（成功率 0.0%，上一代 Argon 为 0.7%）。 网络安全是 Carbon 的头号卖点：在真实漏洞发现基准 SEC-bench Pro 拿下 88.2%，漏洞利用验证 ExploitGym 54.8%、补丁生成 CWE-bench 71.0%，三项全部刷新纪录。内部测试中，它生成安全代码的成功率比两周前的 Argon 高出 2 倍多；Chrome 团队报告其补丁质量是另一家实验室旗舰模型的 10 倍以上。 不过普通用户暂时用不上它：Carbon 本周起先向部分网络防御者开放，之后经 Fairwind 计划逐步扩大到可信伙伴，向开发者和消费者全面开放的日期还没定。价格与 Argon 持平：每百万输入 token 2 美元、输出 10 美元，缓存输入打了 95% 折扣。 > 💬 两周一个旗舰的节奏下，Google 把「不越界」做成了卖点：作弊率归零、注入全防住——安全指标第一次比跑分更像发布会主角。至于普通用户什么时候能用上，官方只说了句「尽快」。
来源：Google 官方博客 / Ars Technica | 2026-10-28
Alphabet 财报夜惊魂：自由现金流转负 90 亿，盘后跌超 10%
Alphabet 在美东周三盘后交出一份让市场倒吸凉气的财报：营收增长 25%，但出现近四年来首次营收与利润双双不及预期：净利润 261.7 亿美元（去年同期 349.8 亿），自由现金流负 90.7 亿美元——远差于分析师预期的负 27 亿。Q3 资本开支 513.6 亿美元，全年指引维持在 1950 亿至 2050 亿，公司并预告 2027 年资本开支将突破 3000 亿。为了填资金缺口，长期债务一年内从 216 亿飙到 1251 亿美元，涨了约 4.8 倍。盘后股价跌超 10%。 > 💬 自由现金流负 90 亿还敢把全年 2000 亿的开支指引稳稳举着，这不是财报，是信心投票——只是投票用的筹码是借来的。
来源：CNBC | 2026-10-28
微软财报的双重彩蛋：Azure 首次晒出 332 亿美元，OpenAI 一家贡献 100 亿
同一天发财报的微软甩出两个「首次」：一是 Azure 季度收入首次以美元披露——332 亿美元，同比增 48%，略高于同季 Google Cloud 的 307 亿（后者增速约为其两倍）；二是监管文件显示，本季度从 OpenAI 相关商业安排（含收入分成）录得 100 亿美元收入，占公司 923 亿季度总收入的近 11%。上一财年全年这个数字是 241 亿，按当前节奏年化约 400 亿。另一面：OpenAI 的应收账款已滚到 72 亿美元，比三个月前又多了 12 亿。 > 💬 一边给 OpenAI 记 100 亿收入，一边账上躺着 72 亿应收款——这段关系里最稳的商业模式，可能是当债主。
来源：TechCrunch | 2026-10-28
Anthropic 上新用量报告工具：企业终于能看清 token 花到哪了
Anthropic 发布面向企业客户的报告工具套件：管理者可以按日期范围查看团队用 AI 干了什么、产出了什么、花了多少钱、流程在哪里卡壳，并附带改进建议；这些数据默认不用于模型训练。招股书里的两个数字给这次发布做了注脚——公司 Q2 已实现运营利润，净收入留存率高达 900%（同一批客户的季度消费一年翻了九倍）。 > 💬 900% 留存率的意思是：客户用得越久，花得越多。现在它把「花到哪了」做成产品——企业买的不再是 token，是能报销的效率。
来源：TechCrunch | 2026-10-28
9.5 分固件漏洞掀桌：EDK2 藏了 16 年的越界写入，牵动全球服务器
安全研究员披露 EDK2（全球绝大多数服务器 UEFI 固件的公共基座）中的一个 9.5/10 级漏洞 CVE-2026-97631：PXE 网络启动驱动处理 ICMP 错误消息时缺少 504 字节的边界检查，可被用来在操作系统和 Secure Boot 等防御加载之前执行未经授权的代码。这个缓冲溢出已经潜伏 16 年。更尴尬的是披露流程：原定明年 1 月的协同披露，被一家主板厂商上个月的代码提交提前掀了盖子。好消息是利用门槛不低——需要对目标所在本地网络的控制权，且纯磁盘启动的家用设备不受影响。 > 💬 16 年的洞、9.5 的分、被一次代码提交掀了盖子。固件是整台机器最底层的地基，而地基的补丁，要靠千万台服务器一台台刷。
来源：Ars Technica | 2026-10-27
乔布斯的规矩作废了：MacBook Pro 第一次配上触屏，还是双层 OLED
苹果发布首款触屏 MacBook Pro——MacBook Pro with PROLED XDR Display，14.2 和 16.3 英寸两种尺寸，采用双层 OLED 面板：2000 尼特峰值亮度、1000 万比 1 对比度、最长 24 小时续航，机身更薄更轻。注意配置分布：触屏款配的是 M5 Pro / M5 Max 芯片，入门款 14 英寸（mini-LED 屏）才是新的 M6，起售价 1999 美元。从周一的 iPad mini 到周三的触屏 MacBook，苹果用三连发把「Mac 不触屏」这条立了二十多年的产品哲学正式送进了历史。 > 💬 乔布斯立了 20 多年的规矩正式作废。Mac 触不触屏从来不是技术问题，是产品哲学问题——现在哲学让位给了现实。
来源：Apple Newsroom / Ars Technica | 2026-10-28
Google 开放 Connected Apps 平台：app 排队等 AI 调度的时代来了
Google 发布 Connected Apps Platform，让第三方开发者把自己的应用接入 Search 和 Gemini 的 AI 体验。此前 connected apps 只限 Google 挑选的大牌服务（用 Canva 设计传单、用 YouTube Music 建歌单），现在企业可以自助接入。逻辑很简单：用户的入口正在从「打开某个 app」变成「在 AI 里说一句话」，谁占住对话框，谁就是新的首页。ChatGPT 和 Meta Muse 已经在做同样的事。 > 💬 入口争夺战的新回合：app 不再求你打开它，而是排队等 AI 调度。分发权从应用商店搬进了对话框。
来源：TechCrunch | 2026-10-28
Nvidia 开源语音转写模型：40 种语言、8 个说话人、单芯片能跑
Nvidia 发布开放权重转写模型 Nemotron 3.5 Transcribe：支持 40 种语言转写、31 种语言与英语互译、一小时长音频、区分 8 个说话人，30B 参数可在单芯片上运行（从 Jetson AGX Thor 到 DGX Spark）。同场开源了语音数据集 Granary v2 和微调配方，号称在 Hugging Face Open ASR 排行榜上领先。瞄准的是会议记录和医疗转写两个成熟市场。一个限制：只处理录音，不支持直播流。 > 💬 转写是 AI 变现最老的场子之一，Nvidia 这次连数据集带微调配方全开源——卖卡的顺手把软件生态也圈了。
来源：TechCrunch / Hugging Face | 2026-10-28