AI Builders Digest - 2026-06-12
X / TWITTER
Box CEO Aaron Levie
Aaron Levie shared Box's early enterprise eval results for Anthropic's Fable model, arguing that the biggest jump is not just coding but "complex knowledge work" over real enterprise documents. In Box AI's eval, Fable beat Opus 4.8 across media, technology, financial services, and healthcare, with the strongest pattern being more consistent multi-step reasoning and fewer shortcut errors.
Aaron Levie 分享了 Box 对 Anthropic Fable 模型的企业场景评测。他的重点不是单纯 coding，而是企业文档里的复杂知识工作：Fable 相比 Opus 4.8 在媒体、科技、金融、医疗等行业任务上都有明显提升，核心差异是多步推理更稳定，不容易走捷径或算错关键步骤。
Source: https://x.com/levie/status/2064922814688481678
Claude / Anthropic
Claude highlighted Cursor cofounder Michael Truell in its Problem Solvers series, noting that Cursor grew from 15 people to 700 in two years and is now used by more than 60% of the Fortune 500. Claude also announced scheduled deployments and vault-backed environment variables for the Claude Platform.
Claude 官方重点推了 Cursor 联合创始人 Michael Truell 的案例：Cursor 两年内从 15 人增长到 700 人，超过 60% 的 Fortune 500 企业在使用它的 AI coding platform。Claude Platform 也上线了 scheduled deployments 和 vault 中的环境变量能力。
Sources: https://x.com/claudeai/status/2064757537992249734, https://x.com/claudeai/status/2064741184547795408
Thariq, Claude Code at Anthropic
Thariq explained how Fable was used to edit its own launch video: the model wrote code and tool calls for transcription, ffmpeg, color grading, Figma MCP access, Remotion UI, and rendering. The notable signal is that video production is becoming an agent workflow, not just a timeline-editing workflow.
Anthropic 的 Thariq 展示了一个很强的 agent-native 创作案例：用 Fable 编辑 Fable 自己的发布视频。模型写了大量代码和 tool calls，完成 transcription、ffmpeg、color grading、Figma MCP、Remotion UI 和渲染。信号很明确：视频生产正在从剪辑软件时间线，转向 agent 工作流。
Source: https://x.com/trq212/status/2064826394589442448
Zara Zhang
Zara Zhang argued that teams should build agents and skills for the cross-functional teams they work with. Her design-team example is sharp: designers can encode brand guidelines and design patterns into an agent so marketing can produce on-brand assets without waiting on design bandwidth. She also warned that too many SF startups are selling to the same narrow audience of AI-native engineering and product teams instead of building for the wider market.
Zara Zhang 提出一个值得记住的组织判断：团队应该为跨职能协作对象构建 agents / skills。比如设计团队可以把品牌规范和设计模式做成 agent，让市场团队更自主地产出符合品牌的素材。她同时提醒，旧金山大量 startup 都在卖给同一小撮 AI-native 工程和产品团队，反而很少有人面向更广泛的 99% 市场。
Sources: https://x.com/zarazhangrui/status/2064835289559023958, https://x.com/zarazhangrui/status/2064825302359150870
Madhu Guru, former Gemini product leader
Madhu Guru offered a practical model-selection rule from early Gemini enterprise work: if you are replacing a known traditional ML system, start small because you already know what good looks like; if you are building something new, start with the most capable model first, discover what is possible, then optimize down to a smaller model later.
前 Gemini 产品负责人 Madhu Guru 给了一个很实用的模型选型规则：如果是在替换已有传统 ML 系统，可以从小模型开始，因为好坏标准已经明确；如果是在做全新应用，应先用最强模型探索可能性，等高质量流程跑通后，再降到更小模型控制成本。
Source: https://x.com/realmadhuguru/status/2064794601320481150
Peter Yang
Peter Yang's strongest point was cultural: give yourself permission to remain a builder. He argues that the old career ladder pushed strong makers into review meetings and management rituals, while AI is making companies reward IC building again. He also noted that the more he uses Codex, the more ambitious his requests become.
Peter Yang 的核心观点是组织文化层面的：允许自己继续做 builder。传统职业阶梯经常把强执行者推向管理、评审和跨部门对齐，但 AI 正在让公司重新奖励 IC 式的构建能力。他还提到，越使用 Codex，越会自然提出更有野心的任务。
Sources: https://x.com/petergyang/status/2064799855059616172, https://x.com/petergyang/status/2064748427892945313
Dan Shipper, Every CEO
Dan Shipper connected AI productivity to reshoring: if each employee can produce more with AI, it becomes more attractive to bring some work back closer to customers in the US. This is a useful macro lens for AI-native operations, not just a productivity story.
Every CEO Dan Shipper 把 AI 生产力提升和岗位回流联系起来：当单个员工借助 AI 的产出显著提升，把部分工作迁回美国、贴近客户就更有经济吸引力。这不是单纯效率叙事，而是 AI-native operations 的宏观结构变化。
Source: https://x.com/danshipper/status/2064777216656097445
Google Labs and Gemini
Google Labs expanded Project Genie access to Google AI Ultra 5X subscribers globally, while Gemini product leader Josh Woodward acknowledged and then resolved a Gemini outage. The product signal is continued packaging of advanced AI experiments behind higher-tier subscriptions, with reliability still a visible part of the user-facing story.
Google Labs 将 Project Genie 扩展给全球 Google AI Ultra 5X 订阅用户；Gemini 负责人 Josh Woodward 则公开确认并修复了一次 Gemini outage。这里的产品信号是：高级 AI 实验能力继续被打包进更高等级订阅，同时稳定性仍然是用户能直接感知的关键体验。
Sources: https://x.com/GoogleLabs/status/2064801929339752527, https://x.com/joshwoodward/status/2064762269674918013
Replit CEO Amjad Masad
Amjad Masad pointed to Replit as a way to automate job search workflows and also highlighted a launch around enterprise agents. The pattern is consistent with Replit's current positioning: consumer-friendly agent workflows on one side, enterprise automation on the other.
Replit CEO Amjad Masad 展示了用 Replit 自动化求职流程，也转发了一个 enterprise agents 相关发布。这个方向符合 Replit 当前的双线定位：一边降低个人用户使用 agent workflow 的门槛，一边切入企业自动化场景。
Sources: https://x.com/amasad/status/2064864439275536495, https://x.com/amasad/status/2064806473352540643
Garry Tan, Y Combinator CEO
Garry Tan highlighted Nessie as a way to move context, memory, and history from ChatGPT, Perplexity, and Gemini into other memory systems, including OpenClaw/Hermes Agent through OpenClaw and MCP servers. The interesting builder signal is that memory portability is becoming infrastructure, not a feature checkbox.
YC CEO Garry Tan 提到 Nessie 可以把 ChatGPT、Perplexity、Gemini 里的 context、memory 和 history 迁移到其他 memory 系统，也能接入 OpenClaw / Hermes Agent。他传递出的 builder 信号是：memory portability 正在从产品功能点，变成 agent infrastructure 的底层能力。
Source: https://x.com/garrytan/status/2064947145652994510
PODCASTS
AI & I by Every - How Anthropic Uses Claude Fable 5 With Mike Krieger
The Takeaway: Mike Krieger's main lesson from using Fable is that strong models force a new work style: less step-by-step prompting, more intent-rich delegation, planning, verification, and parallel long-running sessions.
Krieger, Anthropic Labs head and Instagram cofounder, describes a shift from "ask the model to start" toward giving it a broad goal, enough context, and room to finish the swing. The most concrete change is trust over longer horizons: he can hand off complex work overnight, expect the model to recover from blockers, scaffold missing services, document what happened, and continue once dependencies come back. He still emphasizes review, but the mental model has moved from assistant to teammate.
He also shows what "agent-native software" can mean in practice: a personal media tracker where the embedded agent can not only search and add items, but modify the app itself through managed agents, live previews, diffs, and deployable changes. The deeper point is that the software team is moving closer to the software surface. Users will increasingly expect to ask the app to change itself.
核心判断：Mike Krieger 使用 Fable 的最大体感，不是模型回答更聪明，而是工作方式要变了：少一点逐步 prompt，多一点高意图委托、前置规划、结果验证和并行长任务。
Krieger 是 Anthropic Labs 负责人，也是 Instagram 联合创始人。他描述的变化是，从“让模型开始做某一步”，转向给模型一个足够完整的目标、上下文和完成空间。最具体的变化是长时间尺度上的信任：可以把复杂任务交给模型过夜，让它自己处理 blocker、临时搭后端、记录过程，并在依赖恢复后继续推进。人仍然要 review，但心智模型已经从 assistant 更接近 teammate。
他还展示了一个 agent-native software 的方向：一个个人媒体追踪应用，内置 agent 不只能搜索和添加条目，还能通过 managed agents、live preview、diff 和部署流程直接修改应用本身。更深的信号是，软件团队正在靠近软件界面本身；未来用户会越来越自然地要求应用“自己改自己”。
Source: https://www.youtube.com/watch?v=XWpTgCvgYaE
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders