AI Builders Digest — 2026-06-18
X / TWITTER
Google Labs VP Josh Woodward: Google AI Futures Fund expands to Brazil
Josh Woodward said Google has expanded the Google AI Futures Fund to Brazil through a partnership with Monashees and the new Gama Fund. The offer is clearly aimed at deep tech AI founders: 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 the new IPT Open campus hub.
Josh Woodward 表示，Google 已通过与 Monashees 合作和新的 Gama Fund，把 Google AI Futures Fund 扩展到巴西。这个计划明显面向 deep tech AI 创始人：提前使用 Google DeepMind 模型、最高 200 万美元共同投资、35 万美元 Google Cloud 和 Gemini credits，以及与 Google 工程师在 IPT Open campus hub 直接共创。
Source: https://x.com/joshwoodward/status/2067025851829330076
OpenAI Codex engineer Thibault Sottiaux: Codex features and capacity fixes
Thibault Sottiaux said OpenAI is rolling out major Codex features across Europe and separately acknowledged elevated Codex error rates from "model at capacity." He later said the issue had been fixed and asked users to give the team 24 hours to reset Codex rate limits across all plans.
OpenAI Codex 工程师 Thibault Sottiaux 表示，OpenAI 正在欧洲推出一批重要 Codex 功能；他也承认部分 Codex 用户遇到 "model at capacity" 导致的高错误率。随后他补充说问题已经修复，并请用户给团队 24 小时重置所有套餐的 Codex rate limits。
Sources: https://x.com/thsottiaux/status/2067064381855187231, https://x.com/thsottiaux/status/2066956441173323943, https://x.com/thsottiaux/status/2066865154902380796
Former Google product leader Madhu Guru: Cursor's real asset is the agentic harness
Madhu Guru argued that the real prize in the SpaceX-Cursor deal is not just a coding tool, but Cursor's production-grade agentic harness: planning, context management, tool use, iteration, verification, memory, and error recovery. His view is that this harness can become a general core for automating knowledge work, because Cursor combines model expertise, evals, application design, and GTM around one sharp job: helping software engineers build.
前 Google 产品负责人 Madhu Guru 认为，SpaceX-Cursor 交易真正有价值的不是一个 coding tool，而是 Cursor 的 production-grade agentic harness：规划、上下文管理、工具使用、迭代、验证、记忆和错误恢复。他的判断是，这套 harness 可能成为自动化知识工作的通用核心，因为 Cursor 把模型、evals、应用设计和 GTM 都围绕一个清晰任务打透：帮助软件工程师构建软件。
Source: https://x.com/realmadhuguru/status/2066935654500671499
Vercel CEO Guillermo Rauch: longer functions and sandboxes
Guillermo Rauch highlighted two platform shifts from Vercel: 30-minute function invocations and 24-hour sandbox lifetimes. For AI app builders, this points toward infrastructure that can support longer-running agent tasks, richer background execution, and development environments that survive beyond short request-response windows.
Vercel CEO Guillermo Rauch 提到 Vercel 的两个平台变化：30 分钟 function invocations，以及 24 小时 sandbox lifetimes。对 AI 应用开发者来说，这意味着基础设施正在更适合长时间 agent 任务、更复杂的后台执行，以及不再局限于短请求周期的开发环境。
Source: https://x.com/rauchg/status/2067137678772937000
Box CEO Aaron Levie: open weights and the applied AI layer
Aaron Levie framed one of AI's biggest market questions as the distance between open weights and closed frontier models. If open weights stay only three to six months behind, it changes chip demand, sovereign AI, inference placement, applied AI margins, and how much companies can afford to spend on AI. He also called the Cursor deal symbolically important as the first mega-success template for applied AI: deep domain focus, model routing, frontier model leverage, selective internal training, and strong GTM.
Box CEO Aaron Levie 把 AI 市场的核心问题之一定义为：open weights 模型和闭源 frontier models 到底差多远。如果 open weights 只落后三到六个月，芯片需求、sovereign AI、推理部署位置、应用层利润率、企业 AI 预算都会被改写。他还认为 Cursor 交易具有象征意义，是应用层 AI 的第一个大型成功模板：深领域聚焦、model routing、善用 frontier models、必要时自训练，以及强 GTM。
Sources: https://x.com/levie/status/2067070918300664161, https://x.com/levie/status/2066908002809221496
Cursor designer Ryo Lu: designers can now build the real product
Ryo Lu said the striking thing about Cursor mobile is that a designer coded much of the real thing with Cursor. His takeaway was simple: titles matter less when the toolchain lets more people build production software directly.
Cursor 设计师 Ryo Lu 说，Cursor mobile 最值得注意的一点是，一位设计师用 Cursor 写出了大量真实产品代码。他的核心判断很直接：当工具链让更多人能直接构建 production software，职能标题的重要性会下降。
Source: https://x.com/ryolu_/status/2067124871226929526
Zara Zhang: AI agents need opinion and soul
Zara Zhang pushed back on the flood of generic "AI agent that does everything" products. Her argument: if a product wants to compete with Claude or Codex, it needs a specific opinion, a sharp wedge, and a soul. "Build small & sharp, not big & generic" is the useful product principle here.
Zara Zhang 反对现在大量泛化的 "AI agent that does everything" 产品。她的观点是：如果一个产品想让用户不用 Claude 或 Codex 而用你，就必须有明确观点、锋利切口和自己的 soul。这里最有价值的产品原则是：小而锐，不要大而泛。
Source: https://x.com/zarazhangrui/status/2066936706281206165
FPV Ventures partner Nikunj Kothari: be in the judgment path or token path
Nikunj Kothari said the Cursor acquisition sets a path for more application companies and summarized the strategic placement as: be in the judgment data path or the token path. That is a crisp way to describe where AI application companies can capture durable leverage: own the workflow decisions, or own the model/inference flow.
FPV Ventures partner Nikunj Kothari 认为，Cursor acquisition 会给更多应用公司打开路径。他把战略位置总结为：要么在 judgment data path，要么在 token path。换句话说，AI 应用公司要形成长期杠杆，要么掌握工作流中的判断数据，要么掌握模型和推理流量。
Source: https://x.com/nikunj/status/2066905445974102384
Every CEO Dan Shipper: browser agents still have reliability issues
Dan Shipper said he switched back from Atlas Browser to Dia because of persistent bugs. It is a small product signal, but useful: AI browser products are still being judged first on reliability and daily workflow fit, not just agentic ambition.
Every CEO Dan Shipper 说自己因为 Atlas Browser 持续出现奇怪 bug，已经切回 Dia。这是一个小但有价值的产品信号：AI browser 产品首先仍然会被可靠性和日常工作流适配度检验，而不只是 agentic 愿景。
Source: https://x.com/danshipper/status/2066914130863473048
South Park Commons GP Aditya Agarwal: Snowflake's AI shift
Aditya Agarwal highlighted an upcoming South Park Commons session with Sridhar Ramaswamy, who grew Google's ads business from $1.5B to $100B+ and is now leading Snowflake through the AI shift. The signal is less about the event itself and more about enterprise AI: data infrastructure companies are being forced to turn AI from narrative into operating model.
South Park Commons GP Aditya Agarwal 提到，Sridhar Ramaswamy 将参加 SPC 活动。Sridhar 曾把 Google 广告业务从 15 亿美元做到 1000 亿美元以上，现在正在带领 Snowflake 穿越 AI 转型。这里的信号不只是活动本身，而是 enterprise AI：数据基础设施公司必须把 AI 从叙事变成真实运营模型。
Source: https://x.com/adityaag/status/2066915803610370098
PODCASTS
Training Data: "Simulating Humans at Scale: Simile's Joon Sung Park"
The Takeaway: Joon Sung Park's bet is that the next important AI platform may not be a smarter assistant, but a simulator for human behavior at market and society scale.
Joon Sung Park, founder and CEO of Simile and one of the researchers behind Stanford's "generative agents" work, is turning agent simulations into a company. The origin story starts with Smallville, a simulated town of 25 agents with memory, planning, and reflection, where agents formed routines, relationships, and emergent events like a Valentine's Day party. The company thesis is now much more commercial: enterprises want to answer market and human-insight questions faster than surveys, panels, or field tests allow.
Park's key distinction is that today's frontier models are optimized toward rational superintelligence, while real people are subjective, diverse, inconsistent, and often irrational. That means better benchmarks alone do not make better human simulators. Simile grounds agents in real human data through partners like Gallup, then uses those agents to answer many follow-on questions beyond the original survey. The important phrase is the "say-do gap": models trained on what people say online still need behavioral grounding to predict what people will actually do.
For AI builders, the useful pattern is that simulation may become a new enterprise primitive. Instead of only automating work after decisions are made, AI can help test decisions before they meet the real world.
核心 takeaway：Joon Sung Park 的赌注是，下一类重要 AI 平台可能不是更聪明的 assistant，而是能在市场和社会尺度模拟人类行为的 simulator。
Joon Sung Park 是 Simile 创始人兼 CEO，也是 Stanford "generative agents" 研究背后的核心研究者之一。他正在把 agent simulations 变成公司。起点是 Smallville：一个由 25 个带有记忆、规划和反思能力的 agent 组成的模拟小镇，里面的 agent 会形成日常作息、人际关系，甚至自发组织情人节派对。现在公司的商业命题更明确：企业想比问卷、panel、field test 更快回答市场和人类洞察问题。
Park 的关键区分是：今天的 frontier models 多数朝着理性 superintelligence 优化，但真实人类是主观、多样、不一致、经常非理性的。因此，仅仅提高模型 benchmark 不等于更好的 human simulator。Simile 通过 Gallup 这类合作伙伴采集真实人类数据，把 agent 建立在真实行为基础上，再用这些 agent 回答远超原始问卷范围的后续问题。这里最重要的词是 "say-do gap"：只从网上人们说过的话训练出来的模型，仍然需要行为数据 grounding，才能预测人们真正会做什么。
对 AI builders 来说，值得关注的模式是：simulation 可能成为新的 enterprise primitive。AI 不只是决策后的自动化工具，也可以成为决策进入真实世界前的测试场。
Source: https://www.youtube.com/watch?v=lfhFmwcESRw
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders