June 17, 2026

Google VP Josh Woodward announced that Google is expanding its AI Futures Fund to Brazil, partnering with venture firm Monashees to launch the Gama Fund for an elite cohort of deep-tech founders. The package on offer is unusually generous: early access to Google DeepMind models, up to $2M in co-investment, $350K in Google Cloud and Gemini credits, and direct co-development with Google engineers at a new IPT Open campus hub. https://x.com/joshwoodward/status/2067025851829330076

谷歌副总裁 Josh Woodward 宣布，Google 正把 AI Futures Fund 扩展到巴西，携手风投机构 Monashees 推出 Gama Fund，面向一批精英级 deep-tech 创始人。这套组合相当慷慨：可提前使用 Google DeepMind 的模型、最高 200 万美元的跟投、35 万美元的 Google Cloud 与 Gemini 额度，以及在全新的 IPT Open 校园中心与 Google 工程师直接共同开发。https://x.com/joshwoodward/status/2067025851829330076

OpenAI's Thibault Sottiaux, who leads Codex and ChatGPT, spent the day in firefighting-then-shipping mode. After Codex users hit a wave of "model at capacity" errors, his team fixed the root cause and announced they're resetting Codex rate limits across all plans, teasing that more is on the way. He's also in France for the week, rolling out Codex's newest features across Europe. https://x.com/thsottiaux/status/2066956441173323943 · https://x.com/thsottiaux/status/2067064381855187231

OpenAI 负责 Codex 与 ChatGPT 的 Thibault Sottiaux 这一天先救火、后发货。Codex 用户遭遇一波"model at capacity"报错后，他的团队修好了根因，并宣布将在所有套餐上重置 Codex 的 rate limit，还暗示后面还有动作。他本周人在法国，正把 Codex 最新的一批功能铺到整个欧洲。https://x.com/thsottiaux/status/2066956441173323943 · https://x.com/thsottiaux/status/2067064381855187231

Peter Yang, who publishes practical AI tutorials, is releasing a guide on turning Codex or Claude Code into your personal advisor using a skill built from just four files — a lightweight, reproducible recipe for anyone who wants a custom advisor agent without heavy scaffolding. https://x.com/petergyang/status/2067056979974160749

专注实用 AI 教程的 Peter Yang 即将发布一篇指南，教你用一个仅由四个文件构成的 skill，把 Codex 或 Claude Code 变成你的私人顾问——这是一份轻量、可复现的配方，让任何想要定制顾问 agent 的人都不必搭一堆脚手架。https://x.com/petergyang/status/2067056979974160749

Madhu Guru, a former Google product leader on Gemini, Veo, and Nano Banana, argues the real prize in the SpaceX–Cursor deal isn't the editor — it's the production-grade agentic harness underneath: planning, context management, tool use, iteration, verification, memory, and error recovery. He says that harness can become the core for automating knowledge work at scale, and that Cursor is rare in doing all three of harness engineering, full-stack AI expertise, and end-to-end product focus well at once. https://x.com/realmadhuguru/status/2066935654500671499

曾在 Google 主导 Gemini、Veo、Nano Banana 的产品负责人 Madhu Guru 认为，SpaceX–Cursor 这桩交易真正值钱的不是那个编辑器，而是它底层那套生产级的 agentic harness：规划、context 管理、工具调用、迭代、验证、memory、错误恢复。他说这套 harness 有望成为大规模自动化知识工作的核心，而 Cursor 难得之处在于同时把 harness 工程、全栈 AI 能力和端到端的产品聚焦三件事都做好了。https://x.com/realmadhuguru/status/2066935654500671499

Vercel CEO Guillermo Rauch shipped two infrastructure upgrades: 30-minute function invocations and new 24-hour sandbox lifetimes — both meaningful for long-running agent workloads that used to bump against tight time limits. https://x.com/rauchg/status/2067137678772937000

Vercel CEO Guillermo Rauch 上线了两项基础设施升级：30 分钟的 function 调用时长，以及全新的 24 小时 sandbox 生命周期——这两点对那些过去常常被时间上限卡住的长时运行 agent 任务都很关键。https://x.com/rauchg/status/2067137678772937000

Box CEO Aaron Levie weighed in twice. First, he called the Cursor deal symbolically significant as the first mega-success in AI's applied layer, proving out the playbook of deep domain focus, acting as a model router, knowing when to lean on frontier models versus train your own, and applied-AI GTM — "the first at scale template for how to execute this playbook." Second, he flagged what he sees as one of AI's biggest open questions: how far open-weights models trail closed ones at any moment — 3 months, 6 months, or years. The answer, he says, will determine the chip stack, where inference can run, what sovereign AI looks like, and AI's margin structure — and right now open weights are holding close to frontier capability. https://x.com/levie/status/2066908002809221496 · https://x.com/levie/status/2067070918300664161

Box CEO Aaron Levie 发了两条。其一，他称 Cursor 这桩交易具有象征意义，是 AI 应用层的第一个超级成功案例，验证了一整套打法：深耕垂直领域、充当 model router、清楚何时依赖前沿模型何时自己训练、以及应用层 AI 的 GTM——是"如何执行这套打法的第一个规模化范本"。其二，他指出 AI 当下最大的悬念之一：在任一时刻，开放权重模型究竟落后闭源模型多少——3 个月、6 个月，还是好几年。他说答案将决定芯片栈格局、inference 能在哪里跑、主权 AI 长什么样，以及 AI 的利润结构——而眼下开放权重正紧咬前沿能力。https://x.com/levie/status/2066908002809221496 · https://x.com/levie/status/2067070918300664161

Cursor designer Ryo Lu revealed that Cursor shipped a mobile app — and that most of the real thing was coded in Cursor by someone whose title is "designer." His takeaway cuts past org charts: "titles don't mean shit. you can just build." https://x.com/ryolu_/status/2067124871226929526

Cursor 设计师 Ryo Lu 透露 Cursor 已上线移动端 App——而且这个真家伙的大部分代码，是由一位头衔为"设计师"的人在 Cursor 里写出来的。他的结论一刀切过组织架构图："头衔屁都不是。你直接动手做就行。"https://x.com/ryolu_/status/2067124871226929526

Builder Zara Zhang offered a sharp contrarian take on AI products: nearly every new product right now pitches itself as "an AI agent that does everything in your work and life and integrates with everything" — but that's just Claude or Codex. To get someone to use your thing instead, it needs an opinion and a soul: build small and sharp, not big and generic, because "doing everything means doing nothing." Her related advice: don't chase the cool thing — go so deep on something everyone else thinks is too small that it becomes the cool thing. https://x.com/zarazhangrui/status/2066936706281206165 · https://x.com/zarazhangrui/status/2066994434953421226

Builder Zara Zhang 对 AI 产品给出了一个犀利的反主流判断：如今几乎每个新产品都把自己包装成"一个能搞定你工作和生活里一切、并能和一切打通的 AI agent"——可那不就是 Claude 或 Codex 嘛。想让人改用你的东西，它得有观点、有灵魂：做小而锋利的，别做大而通用的，因为"什么都做，就等于什么都没做"。她另一条相关建议是：别去追那个酷东西——把一件所有人都觉得太小的事钻得足够深，它自己就会变成那个酷东西。https://x.com/zarazhangrui/status/2066936706281206165 · https://x.com/zarazhangrui/status/2066994434953421226

FPV Ventures partner Nikunj Kothari distilled the Cursor acquisition into a positioning rule for startups: be in the judgement (data) path or the token path. He reads the deal as setting the template for more application-layer companies to follow. https://x.com/nikunj/status/2066905445974102384

FPV Ventures 合伙人 Nikunj Kothari 把 Cursor 这桩收购提炼成一条给创业公司的定位法则：要么站在判断（数据）链路上，要么站在 token 链路上。他认为这桩交易为更多应用层公司立下了可以照着走的范本。https://x.com/nikunj/status/2066905445974102384

Every CEO Dan Shipper shared a candid product signal: he switched off OpenAI's Atlas Browser and back onto Dia, citing a string of weird bugs and no sense that Atlas is improving. https://x.com/danshipper/status/2066914130863473048

Every CEO Dan Shipper 给出了一个直白的产品信号：他已经弃用 OpenAI 的 Atlas Browser，转回 Dia，理由是 Atlas 一连串的诡异 bug，而且看不出它在变好。https://x.com/danshipper/status/2066914130863473048

Building intelligent apps for Apple platforms with Claude in the Foundation Models framework announces a new Swift package that lets Apple developers call Claude through Apple's own Foundation Models framework. The framework already exposes Apple's on-device models to Swift for fast local tasks like summarization and extraction, returning typed Swift values via guided generation in as few as three lines of code. The package extends that pattern: when a request needs multi-step reasoning, code generation, web search for current info, or code execution for data analysis, developers can hand off to Claude and stream the response back into the same SwiftUI view. The key ergonomic win is that because Apple's framework returns typed values from `@Generable` annotations, "developers arrive at the Claude API call with clean inputs instead of raw user text." The post sketches concrete patterns — a journaling app that generates daily prompts on-device, then asks Claude to find threads across months of entries; a study app that defines a term locally, then hands off to Claude for "why does this matter for everything else?" follow-ups. It works across iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27: add the package, sign in with an Anthropic API key, and pass typed outputs into a Claude request, with the package handling streaming, tool calls, and structured responses. https://claude.com/blog/claude-for-foundation-models

Claude 博客的文章《Building intelligent apps for Apple platforms with Claude in the Foundation Models framework》宣布推出一个新的 Swift package，让 Apple 开发者可以通过 Apple 自家的 Foundation Models framework 来调用 Claude。该 framework 本就把 Apple 的端侧模型开放给 Swift，用于摘要、抽取这类快速的本地任务，并通过 guided generation 用最少三行代码返回带类型的 Swift 值。这个 package 把这一模式向外延伸：当请求需要多步推理、代码生成、为获取实时信息而联网搜索，或为数据分析而执行代码时，开发者可以交接给 Claude，并把回复 stream 回同一个 SwiftUI 视图。最关键的体验提升在于：由于 Apple 的 framework 会从 `@Generable` 注解返回带类型的值，"开发者抵达 Claude API 调用时拿到的是干净的输入，而不是原始的用户文本"。文中给出了具体范式——一个日记 App 在端侧生成每日提示，再让 Claude 在数月的条目中找出线索；一个学习 App 在本地给出术语定义，再交接给 Claude 回答"这为什么对其他一切都重要？"这类追问。它支持 iOS 27、iPadOS 27、macOS 27、visionOS 27 和 watchOS 27：装上 package、用 Anthropic API key 登录，把带类型的输出传进一次 Claude 请求即可，streaming、工具调用和结构化返回都由 package 处理。https://claude.com/blog/claude-for-foundation-models

The Takeaway: The race to build ever-more-rational superintelligence may be the wrong frontier for understanding people — because humans are gloriously irrational, and capturing that requires a fundamentally different kind of model.

Joon Sung Park is the founder and CEO of Simile, an applied AI lab simulating human behavior and societies. He's the Stanford researcher behind "Smallville," the 2023 experiment where 25 generative agents lived in a virtual town — waking up, going to work, forming relationships, and even throwing an emergent Valentine's Day party. His through-line since a 2020 foundation-models paper has been one bet: large language models encode human behavior, and you can build simulations on top of it.

The contrarian core is this. Today's frontier labs aim their north star at rational, superintelligent machines that excel at problems with objective answers. But "people are irrational. We have a lot of subjective values, preferences and tastes. So you actually start to see divergence in model size going up and the performance in its ability to predict and simulate human behavior." Park's framing: if today's models are the "CPU of intelligence" — single, rational, brilliant at objective problems — Simile is building the "GPU of intelligence," many subunits that faithfully represent the diversity of real subpopulations. Not superhuman; as human as possible.

To close the say-do gap, Simile grounds simulations in real behavioral data and even life-story interviews, then validates rigorously — its US-population study predicted people's behavior 85% as accurately as those people replicate themselves, with a total-variation-distance threshold under 0.15 for decision-grade use. Customers like CVS already use it for concept testing and even simulating earnings calls. The bigger dream is Nobel-scale: simulating segregation, climate collective action, or the warning signs of a democracy about to collapse — a Hubble telescope for human society.

核心要点： 一味追逐越来越"理性"的超级智能，也许是理解人类的错误方向——因为人本就酣畅淋漓地不理性，而要捕捉这一点，需要一种根本不同的模型。

Joon Sung Park 是 Simile 的创始人兼 CEO，这是一家模拟人类行为与社会的应用型 AI 实验室。他正是 "Smallville" 背后的斯坦福研究者——在那个 2023 年的实验里，25 个生成式 agent 生活在一座虚拟小镇上：早上醒来、去上班、建立关系，甚至自发地办起了一场情人节派对。自 2020 年那篇基础模型论文以来，他始终押注一件事：大语言模型编码了人类行为，你可以在其之上搭建模拟。

反主流的内核在于：今天的前沿实验室把北极星定在理性的、擅长求解有客观答案问题的超级智能机器上。但"人是不理性的。我们有大量主观的价值观、偏好和品味。于是你会看到，模型规模越大，它预测和模拟人类行为的能力反而开始背离"。Park 的比喻是：如果说今天的模型是"智能的 CPU"——单一、理性、在客观问题上无比出色——那 Simile 要造的是"智能的 GPU"，由许多忠实代表真实子人群多样性的子单元组成。不求超人，但求尽可能像人。

为弥合"说与做"的鸿沟，Simile 把模拟扎根于真实的行为数据，甚至是讲述人生故事的访谈，然后严格验证——其针对美国人口的研究，对人们行为的预测准确度达到了这些人自我复现水平的 85%，并把可用于决策的总变差距离（TVD）门槛定在 0.15 以下。CVS 等客户已经用它做概念测试，甚至模拟财报电话会。而更大的梦想是诺奖量级的：模拟种族隔离的成因、气候问题上的集体行动困境，或一个民主政体即将崩塌前的预警信号——为人类社会造一台哈勃望远镜。