Daily ·
Anthropic prices its IPO, Meituan ships a 1.6-trillion-parameter model
Anthropic sets its IPO at a US$1.4 trillion valuation, Meituan releases a 1.6-trillion-parameter agent model trained entirely on Chinese chips, Claude Fable 5.2 paints with code, and a Vulcan rocket loses 40 Amazon satellites.
Nine signal stories from the last twenty-four hours, from the pricing of the largest AI listing yet to a rocket failure that cost Amazon forty satellites.
1. Anthropic prices its IPO at a US$1.4 trillion valuation
Anthropic set a price range last week of US$70 to US$78 per share, with the stock — ticker "ANT" — expected to begin trading on Thursday. At the top of that range the offering would raise up to US$64.4 billion and value the company at about US$1.4 trillion including employee stock options and restricted shares, well below the US$2 trillion figure floated earlier in the year. Chief executive Dario Amodei spent Friday pitching investors at a JPMorgan event alongside Jamie Dimon, one of the offering's lead underwriters, arguing that the company's enterprise focus was "synergistic with our focus on safety." Taipei Times · Nov 16
2. Meituan releases a 1.6-trillion-parameter model trained on Chinese chips
Meituan's LongCat team published the weights for LongCat-2.5, a mixture-of-experts model with 1.6 trillion total parameters — roughly 48 billion active per token — built for long, multi-step agent work such as coding. It handles a one-million-token context through the API, accepts text, images and video, and converts images into tokens rather than bolting on a separate encoder. Meituan's self-hosting guide describes a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series and makes no mention of Nvidia GPUs; quantizing weights and activations to 4-bit precision raised prompt throughput 43.3 percent and response generation 42.77 percent over an 8-bit version. The model scores 88.4 on Terminal Bench. The Decoder · Nov 15
3. Claude Fable 5.2 paints with code instead of generating pixels
Anthropic's Claude Fable 5.2 can produce photorealistic paintings without a diffusion model: the language model writes Python programs of 1,000 to 2,000 lines that drive a simulated paint engine, placing hundreds of thousands of individual brushstrokes complete with bristle texture and paint application. An image of a shopping cart in a parking garage took roughly 700,000 strokes, according to Anthropic employee Jake Eaton, who says the model works without an external image model and without pre-rendered or traced source images. A 21-second video of a bird at sunrise uses a counter to track a painting built from exactly 111 strokes, and the model repaints its own mistakes between passes. The Decoder · Nov 15
4. A Vulcan rocket fails and destroys 40 Amazon Leo satellites
United Launch Alliance suffered what is reportedly its first mission failure in its 20-year history when a Vulcan vehicle launched from Space Launch Complex 41 flew into an anomaly ten minutes into flight and failed to deliver its payload of 40 Amazon Leo satellites to orbit. The flight was supposed to be the first of 38 that Amazon has purchased from ULA for Leo, its Starlink competitor, which has so far put only a small portion of the satellites it needs into orbit. ULA said the cause is not yet known; Amazon said ULA "has been a trusted partner" and that it believes the company "will learn from this mission." The Verge · Nov 15
5. Apple's first smart glasses take shape ahead of WWDC
Apple's first smart glasses — prototypes dubbed N50 — are set to be announced at June's WWDC and will lack augmented reality features, according to Bloomberg's Mark Gurman, with shipments "before the end of the year." The prototypes run a pared-down version of visionOS and a new processor based on the company's older smartwatch chips, and use a flat, rectangular MagSafe-like charger Gurman describes as roughly the size of a Tic Tac that can be used while the glasses are being worn. Wi-Fi lets some functions work when the iPhone isn't nearby, though the eyewear won't operate as a fully independent device. The Verge · Nov 15
6. Bharat Semi raises US$24 million for AI-enabled image sensors
Indian semiconductor startup Bharat Semi has raised US$24 million in an all-equity seed round led by Info Edge Group, South Park Commons and Indus Bridge Ventures, to develop programmable image sensors that combine capture, processing and AI on a single chip. The nearly five-year-old, pre-revenue company says it has design wins with US defense technology firm General Atomics and India's Kaynes Semicon, and that collapsing three components into one chip could cut system costs by at least 70 percent. Its image sensors remain under development, with initial commercial production targeted for 2027. TechCrunch · Nov 15
7. Bland launches Tofu-1, a full-duplex voice model for phone calls
Bland released Tofu-1, a full-duplex speech-to-speech model that hears, reasons and replies in audio during live calls, the first in a planned family of models. Most voice agents run as cascades — speech to text, text to a language model, then back to speech — but Tofu-1 listens to audio directly, keeping tone, hesitation and background sound available, and lets callers interrupt or correct it mid-sentence. It can also generate call sound itself, including a second speaker, the room around them and line noise. Bland has not published the internal architecture; research preview access is limited to a small number of enterprises, with banking, healthcare, insurance and government named as target sectors. MarktechPost · Nov 15
8. Singapore's game studios are being squeezed
Ubisoft plans to cut 70 jobs at its Singapore studio — around 20 percent of its local workforce — part of a broader global contraction that has already seen the company close studios and cancel titles, including shutting its RedLynx studio in Finland in the same week. Local developers and observers told CNA that high operating costs, tighter funding and limited job opportunities are squeezing a sector that received fresh government support only two years ago, and pointed to Electronic Arts' closure of BioWare after more than three decades as evidence the problem is industry-wide rather than local. CNA · Nov 16
9. Flaming Hydra's Phoenix Project takes aim at AI clones and licensed e-books
Flaming Hydra, the online culture and commentary magazine, has launched a Kickstarter for The Phoenix Project: a one-off print magazine, an e-book service aimed at libraries, and a newsletter subscription and archiving programme also designed for libraries. Lead editor Maria Bustillos frames the three as one effort to build a "human internet" — an ecosystem that doesn't allow others to create and sell AI-generated clones of writers' work. The pitch follows reporting on Flaming Hydra about an author who found countless AI-generated clones of his book on Amazon. TechCrunch · Nov 15
Try This Next
- Check the LongCat-2.5 self-hosting guide — a working recipe for a 1.6-trillion-parameter agent model on 194 Ascend accelerators is the most useful artefact of the week for anyone planning around export controls. The Decoder
- Watch the Fable 5.2 brushstroke video — 21 seconds, 111 strokes, and every one of them a coordinate in a Python program. It is a genuinely different answer to "how do you make an image." The Decoder
- Save the Phoenix Project as a template — owned e-books and library lending are a concrete anti-clone distribution model, not just a slogan. TechCrunch
每日简报:Anthropic 为 IPO 定价,美团发布 1.6 万亿参数模型
Anthropic 以约 1.4 万亿美元估值启动 IPO,美团发布全程训练于国产芯片的 1.6 万亿参数智能体模型,Claude Fable 5.2 用代码作画,Vulcan 火箭失败损失 40 颗亚马逊卫星。
过去二十四小时里的九条信号,从迄今最大的一笔 AI 上市定价,到一次让亚马逊损失四十颗卫星的火箭失败。
一、Anthropic 为 IPO 定价,估值约 1.4 万亿美元
Anthropic 上周设定每股 70 至 78 美元的发行价区间,股票代码为“ANT”,预计周四开始交易。按区间上限计算,此次发行最多可募资 644 亿美元,包含员工股票期权与限制性股票在内的估值约为 1.4 万亿美元,远低于今年早些时候传出的 2 万亿美元。首席执行官 Dario Amodei 上周五在摩根大通的一场活动上向投资者推介,与承销商之一的 Jamie Dimon 同台,他表示公司聚焦企业客户,与对安全的重视“具有协同效应”。 台北时报 · 11 月 16 日
二、美团发布 1.6 万亿参数模型,全程训练于国产芯片
美团 LongCat 团队公开了 LongCat-2.5 的权重。这是一个混合专家模型,总参数 1.6 万亿,每个 token 约激活 480 亿,面向长时间、多步骤的智能体任务,尤其是编程。通过 API 可一次处理一百万 token 的上下文,接受文本、图像和视频输入,并把图像转换为 token,而不是另外挂接一个编码器。美团的自托管指南描述了一个由 194 块华为昇腾 Atlas A2 加速卡组成的集群,全文未提及 Nvidia GPU;把专家权重和激活量化为 4 位精度后,提示处理吞吐提升 43.3%,回复生成提升 42.77%。该模型在 Terminal Bench 上得分 88.4。 The Decoder · 11 月 15 日
三、Claude Fable 5.2 用代码“作画”,而不是生成像素
Anthropic 的 Claude Fable 5.2 可以在没有扩散模型的情况下画出逼真的画作:语言模型编写 1000 到 2000 行的 Python 程序,驱动一个模拟绘画引擎,落下数十万笔独立的笔触,连笔毛的质感和颜料附着方式都包含在内。据 Anthropic 员工 Jake Eaton 介绍,一幅停车场里购物车的图像大约用了 70 万笔;模型不依赖外部图像模型,也没有使用预渲染或描摹的源图。一段 21 秒的日出之鸟视频用计数器记录一幅恰好由 111 笔构成的画,模型会在每一轮之间重新审视画布,并覆盖自己的失误。 The Decoder · 11 月 15 日
四、Vulcan 火箭发射失败,40 颗亚马逊 Leo 卫星损毁
联合发射联盟(ULA)遭遇据称是其 20 年历史上首次任务失败:一枚从 41 号发射台升空的 Vulcan 火箭在起飞十分钟后出现异常,未能把 40 颗亚马逊 Leo 卫星送入轨道。这次发射本应是亚马逊向 ULA 采购的 38 次飞行中的第一次。Leo 是亚马逊对标 Starlink 的项目,目前入轨卫星数量距离挑战 Starlink 还差很远。ULA 表示故障原因尚不清楚;亚马逊表示 ULA“一直是值得信赖的伙伴”,并相信公司会从这次任务中吸取教训。 The Verge · 11 月 15 日
五、苹果首款智能眼镜成形,将在 WWDC 亮相
据彭博社 Mark Gurman 报道,苹果首款智能眼镜——代号 N50——将在明年 6 月的 WWDC 上发布,且不具备增强现实功能,出货时间在“明年年底之前”。原型机运行精简版 visionOS 和一颗基于旧款手表芯片的新处理器,使用一个扁平矩形、类似 MagSafe 的充电器,Gurman 形容其“约一粒 Tic Tac 大小”,戴在脸上时也能充电。Wi-Fi 让部分功能在 iPhone 不在身边时仍可使用,但它不会成为完全独立的设备。 The Verge · 11 月 15 日
六、Bharat Semi 融资 2400 万美元,做 AI 图像传感器
印度半导体初创公司 Bharat Semi 完成 2400 万美元全股权种子轮,由 Info Edge Group、South Park Commons 和 Indus Bridge Ventures 领投,用于开发把图像采集、处理和 AI 集成在单颗芯片上的可编程图像传感器。这家成立近五年、尚无营收的公司表示已获得美国国防科技公司 General Atomics 和印度 Kaynes Semicon 的设计订单,把三个部件合并到一颗芯片上可使系统成本降低至少 70%。其传感器仍在开发中,首批商业化量产目标定在 2027 年。 TechCrunch · 11 月 15 日
七、Bland 发布 Tofu-1:面向电话通话的全双工语音模型
Bland 发布 Tofu-1,一个在通话中直接听、推理并用音频作答的全双工语音到语音模型,也是其规划中模型家族的首款产品。多数语音智能体采用级联架构——语音转文字、交给语言模型、再转回语音——而 Tofu-1 直接处理音频,保留了语调、犹豫和背景声,并允许来电者中途打断或纠正。它还能生成通话环境音,包括第二个说话人、房间声和线路噪声。Bland 未公开内部架构;研究预览名额仅向少数企业开放,目标行业为银行、医疗、保险和政府。 MarktechPost · 11 月 15 日
八、新加坡游戏工作室承压
育碧计划在新加坡工作室裁减 70 个岗位,约占其本地员工的两成,这是公司全球收缩的一部分——它此前已在多地关闭工作室并取消项目,同一周还关闭了芬兰的 RedLynx 工作室。本地开发者和观察人士对 CNA 表示,高昂的运营成本、收紧的融资和有限的就业机会,正在挤压这个两年前才获得新一轮政府支持的行业;他们并指出艺电关闭运营三十余年的 BioWare,说明问题是全行业性的,而非本地独有。 CNA · 11 月 16 日
九、Flaming Hydra 的 Phoenix 计划:对抗 AI 克隆与受限电子书
网络文化与评论杂志 Flaming Hydra 为“Phoenix 计划”发起众筹:一本一次性印刷杂志、一项面向图书馆的电子书服务,以及同样面向图书馆的通讯订阅与存档项目。主编 Maria Bustillos 把这三件事视为同一个目标——构建一个“人类的互联网”,一个不允许他人生成并销售作者作品 AI 克隆的生态。这一构想源于该刊此前的报道:一位作者在亚马逊上发现了自己书籍的大量 AI 生成克隆。 TechCrunch · 11 月 15 日
值得一试
- 翻一遍 LongCat-2.5 的自托管指南 —— 在 194 块昇腾加速卡上跑 1.6 万亿参数智能体模型的可用配方,是本周对任何围绕出口管制做规划的人最有用的材料。 The Decoder
- 看一遍 Fable 5.2 的笔触视频 —— 21 秒、111 笔,每一笔都是 Python 程序里的一个坐标。
- 把 Phoenix 计划存成模板 —— 可拥有的电子书和图书馆借阅,是一种具体的反克隆分发模式。 TechCrunch
More Dailies
Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.
