Marisa Wu · AI Strategy

On April 15th, OpenAI updated its Agents SDK with sandboxing, configurable memory, and self-correction capabilities. Gartner forecasts 40% of enterprise apps will embed AI agents by end of 2026. AI is no longer just executing tasks — it's entering the decision layer. The question is no longer “which employees will AI replace?” It's “how deeply can AI participate in a CEO's judgment?” That's a diff...

4月15日，OpenAI更新了Agents SDK，加入沙盒运行、自主记忆、自我纠偏能力。Gartner预测2026年底40%的企业应用将内嵌AI Agent。AI正在从“执行工具”变成“决策层”。问题不再是“AI能替代哪些员工”，而是“CEO的判断力，AI能参与多深”。这条线，比任何一次模型发布都更值得认真看。 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit ngmarisa.substack.com

Apple plans to open Siri to Google Gemini and Anthropic Claude at WWDC in June, letting users choose their AI engine. This isn't OpenAI getting kicked out — it's Apple turning Siri from a single-model interface into a multi-model platform. With 2 billion devices at stake, Apple just changed the rules of the AI access war. The question is no longer "which AI is smartest." It's "who controls the ent...

Apple计划在6月WWDC上宣布Siri接入Google Gemini和Anthropic Claude，用户可以自行选择AI引擎。目前ChatGPT已可通过Siri访问，但体验不流畅。这一步不是OpenAI"被踢走"，而是Apple把Siri从"单一AI入口"变成"多模型调度平台"。20亿台设备的AI入口之战，规则刚刚改变。 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit ngmarisa.substack.com

大家好，我是Marisa。 4月7日，Anthropic发布了一个公告，内容大意是：我们造出了一个新模型，但我们不打算把它发布给所有人用。 这不是谦虚，也不是营销。 它叫Claude Mythos Preview。而Anthropic不发布它的原因，是这个模型在测试中做到了一件让公司高层真正感到不安的事。 先说它做了什么。 Anthropic的安全团队让Mythos去测试各种主流软件的安全性。结果： 它在所有主流操作系统和所有主流浏览器中，找到了数千个高危零日漏洞...

Anthropic是AI圈最强调"我们最安全"的公司，连美国国防部都被他们拒过。3月31日，他们意外把Claude Code的512,000行源代码发布到了公网，被fork 41,500次，永久收不回来。原因是工程师打包时忘记排除一个调试文件。这件事说明了一件事：安全文化和安全执行，是两回事。 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit ngmarisa.substack.com

Anthropic — the AI company that refused the Pentagon on safety grounds — accidentally published 512,000 lines of Claude Code's source code to the public internet. A forgotten .map file. 26 million X views. 41,500+ GitHub forks. Permanent. The incident draws a clear line between having a safety culture and actually executing on safety in every engineering detail. Those are not the same thing. This...

SoftBank secured a $40 billion unsecured bridge loan to fund its OpenAI bet — due in 12 months. Five major banks lent the money with no collateral. The hidden assumption: OpenAI IPOs this year. When it does, SoftBank's 13% stake covers the debt. Masayoshi Son's total bet on OpenAI will reach $64.6 billion. This isn't just investing. It's a bet with a deadline. This is a public episode. If you woul...

孙正义借了400亿美元押注OpenAI，无担保，12个月到期。五家顶级银行愿意借，背后只有一个逻辑：相信OpenAI今年IPO。完成后他在OpenAI累计押注646亿，持股13%。这不是投资，这是一场有期限的生死赌局。 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit ngmarisa.substack.com

Microsoft owns 27% of OpenAI — then built a feature where Claude reviews GPT's answers. Not a betrayal. This is platform thinking: Copilot must be the best enterprise research tool, regardless of which model powers it underneath. Whoever integrates multiple models most effectively wins the enterprise market. This is a public episode. If you would like to discuss this with other subscribers or get...

微软持有OpenAI 27%股份，却在自己的旗舰产品里让Claude来审查GPT的答案。这不是背刺，是平台公司的真实逻辑——Copilot必须是最好的企业研究工具，至于底层用谁的模型，让市场决定。 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit ngmarisa.substack.com

OpenAI shut down Sora less than six months after launch, ending Disney's planned $1 billion investment before it even transferred. The official reason: compute costs. The real reason: an IPO is coming, and investors need a clear business model — not experiments. Same week they're hiring 3,500 new people. This is a deliberate strategic consolidation — and Disney is already looking for a new AI part...

OpenAI关闭了Sora，Disney计划中的$10亿投资随之取消。上线不到半年，OpenAI最出名的消费产品就这样结束了。官方原因是算力成本，真实原因是IPO在即，投资人要看清晰的商业模式。一边招人翻倍，一边砍掉最出名的产品——这是一个非常清醒的战略收缩。 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit ngmarisa.substack.com

OpenAI is doubling its headcount to 8,000 by end of 2026. The same week, Microsoft announced it's building its own frontier models, and Google opened Gemini's personal intelligence features to all users for free. Three companies moving simultaneously — in completely different directions. This stopped being about model benchmarks a long time ago. This is a public episode. If you would like to discu...

OpenAI今年要把员工从4500人扩到8000人。但这3500个新人不是去做研究的。同一周，微软开始自研模型拉开距离，谷歌把Gemini数据整合全面免费开放。三家公司同时在移动，这场竞争的维度已经彻底变了。 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit ngmarisa.substack.com

In 2026, five of the world's biggest tech companies — Meta, Amazon, Microsoft, Alphabet, and Oracle — are collectively spending $720 billion on AI infrastructure. Wall Street is cheering. But when you follow the money, the company collecting on every single dollar isn't OpenAI or Google. In this episode, I trace the actual flow of capital, unpack the one supply chain risk nobody wants to say out l...

2026年，全球五大科技巨头——Meta、Amazon、Microsoft、Alphabet、Oracle——今年AI资本支出加总高达$7200亿美元。华尔街掌声不断，但这笔钱最后流进了谁的口袋？不是OpenAI，不是任何一个AI模型公司。这期播客，我们追着钱走，拆开这场史上最大AI军备赛背后的三层真相：谁是最确定的赢家、这条链条上最少人说的那个风险、以及这场投入的回报逻辑为什么已经开始验证。 This is a public episode. If you would like to discuss this w...

大家好，我是Marisa。 昨天是GTC——英伟达一年一度最重要的发布会。我看完了整场。2.5个小时，30,000个人在San Jose的SAP Center里。 [1] 看完之后，有几个地方让我停下来想了很久。 今天这期播客，我不想只复盘“发布了什么”。这些你到处都能看到。我想聊的是：黄仁勋在台上说的这些话，背后的钱在往哪里流。 追钱，不追话术——这一直是我看科技大厂的方法。CEO说什么不重要，钱流到哪里才是真相。 好，先说那句让我后背发凉的话。...

你好，我是Marisa。 今天我想聊一件事，不是芯片，是权力。 大后天，3月16日，黄仁勋要站在圣何塞SAP Center的舞台上，对着从190个国家飞来的3万人发表演讲。所有科技媒体大后天都会报道他说了什么新芯片、什么新架构。但我今天想聊的，是这件事背后一个更大的结构问题—— 为什么每年3月，全球AI圈都要来圣何塞朝圣？ ──────────────────────────────────────────────────────────── 【第一部分：GTC到底是什么】 我先给你一个数字...

3 月 10 日，前 Meta 首席 AI 科学家 Yann LeCun 的新公司 AMI Labs 完成约 10.3 亿美元种子轮融资，估值约 35 亿美金，赌的却不是下一个 ChatGPT，而是能理解物理世界和因果结构的 world models。 这笔钱标记了 AI 赌局的一个拐点：长线资本开始从‘谁的 LLM 更会聊天’，转向‘谁先占领现实世界操作系统的底层架构’。 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus...

Qwen 刚在国际基准上打出「中国模型能和 GPT 对线」的声量，技术扛旗人却在 48 小时内发出「me stepping down. bye my beloved qwen」离场。 这篇不是吃瓜某个大牛的个性，而是从资本和基础设施视角，拆给阿里和所有在做基础模型的大厂看：当你终于有了一张国家级的 AI 底牌，旧式大厂组织本能如何亲手削弱自己的筹码。 This is a public episode. If you would like to discuss this with other subscribers or get access to bon...

OpenAI 拿下 1100 亿融资、估值冲向 8000 亿美金的同一周，黄仁勋在台上淡淡说了一句：这可能是英伟达最后一次投它。这一集，我们顺着这句话往下拆：1000 亿传闻为什么缩成 300 亿现实，这对 AI 赌局的筹码分配意味着什么，对中国大厂技术人和产业链里的普通打工人，各自会改写哪条职业路径。 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit ngmari...

英伟达和全球运营商在押注一件事：未来 6G 本身就是为 AI 和机器而建的公路；与此同时，AI 正在吃光世界内存，让手机变贵，也被 CEO 用来重做组织结构。这期我想聊的，是：在这些巨额赌注里，打工人到底该站哪一层。 This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit ngmarisa.substack.com

2026 年，Google、微软、亚马逊、Meta 准备为 AI 基础设施掏出 6500 亿美金，Nvidia 的业绩飙升，市场却开始担心这一轮投入已经从「技术革命」变成「资金游戏」。这篇文章试图用一篇的篇幅，把钱的去向、各家在赌的未来，以及这场豪赌对工程师、投资者和普通用户各自意味着什么，讲清楚。读完，你至少能回答一个问题：在这盘棋里，你究竟在下棋，还是被当成筹码。 This is a public episode. If you would like to discuss this wi...

TechCrunch 2 月 5 日报道，Amazon 宣布 2026 年计划投入 2000 亿美元资本开支，Google 紧随其后给出 1750–1850 亿美元指引，两家巨头在 AI 基础设施上的投入，已经进入"比谁砸得更狠"的新阶段。 这一期我不比较谁的模型更强，只从三个商业维度拆开看：这两笔钱分别砸在哪些资产上、背后对应的是什么样的收入结构和回报预期，以及当两家巨头同时把 AI 基建当成"护城河"来铺时，站在产业链不同位置的玩家，会被推向怎样的选择。 Thi...