AI Builders Digest — 2026-07-15
X / TWITTER
Swyx
Swyx described his current stack for "Big Boy projects": use Sol Ultra for planning, Fable 5 for critique, Sonnet 5 / Terra Ultra / SWE 1.7 for heavy coding, and Devin Review for review through Kakuna. The useful pattern is not the model list itself, but the workflow: elicit decisions upfront with "grill-me" or "interview-me" style prompts before letting coding agents run.
Source: https://x.com/swyx/status/2076811977918484795
Swyx 分享了他当前处理 "Big Boy projects" 的模型组合：Sol Ultra 负责规划，Fable 5 负责批判，Sonnet 5 / Terra Ultra / SWE 1.7 负责重度编码，Devin Review 通过 Kakuna 做代码审查。真正有价值的不是模型名单，而是工作流：先用 "grill-me" 或 "interview-me" 这类 prompt 把关键决策问清楚，再让 coding agent 执行。
Cat Wu and Thariq, Anthropic
Cat Wu announced an upgrade to Claude Artifacts, while Thariq highlighted that the new capabilities make artifacts more expressive and collaborative. His example: a project dashboard in Claude Tag that can be edited by teammates or by local Claude Code sessions.
Sources: https://x.com/_catwu/status/2076867882894684314 / https://x.com/trq212/status/2076790799011131735
Cat Wu 宣布 Claude Artifacts 获得升级，Thariq 则强调新能力让 artifacts 更具表达力，也更适合协作。他给出的例子是：在 Claude Tag 里创建项目 dashboard，团队成员和本地 Claude Code session 都可以继续编辑。
Amjad Masad, Replit CEO
Amjad Masad showed realtime progress updates for model training runs and framed the experience as feeling like early vibe coding, except now it is making personal models. This is a small but important signal: the same interactive, feedback-rich loop that made coding agents feel magical is starting to move into model creation and customization.
Source: https://x.com/amasad/status/2076776737074184661
Replit CEO Amjad Masad 展示了模型训练任务的实时进度更新，并说这种体验像早期 vibe coding，只不过现在生成的是个人模型。这是一个小但重要的信号：让 coding agent 显得神奇的那种实时反馈循环，正在进入模型创建和定制环节。
Guillermo Rauch, Vercel CEO
Guillermo Rauch said the two most popular features of Vercel's AI-facing product surface are ease of use through a filesystem API and observability, and Vercel is doubling down on both. He also pointed to feature flags as a building block for autonomous, self-optimizing websites where agents can set up and tune experiments.
Sources: https://x.com/rauchg/status/2076817174073880957 / https://x.com/rauchg/status/2076786138195595704
Vercel CEO Guillermo Rauch 表示，Vercel 面向 AI 的产品能力里最受欢迎的是两点：基于 filesystem API 的易用性，以及 observability，接下来会继续加码。他还指出 feature flags 会成为自主优化网站和应用的基础设施，让 agent 可以自行设置和调优实验。
Aaron Levie, Box CEO
Aaron Levie argued that AI will not be a zero-sum stack. Frontier labs keep pushing intelligence forward, open weights absorb breakthroughs and lower costs, applied AI companies win by routing work across models with domain evals and workflow context, and enterprises mostly win by connecting AI to the right data rather than training one model per company.
Source: https://x.com/levie/status/2076882332821373381
Box CEO Aaron Levie 的判断是：AI stack 不会是零和竞争。Frontier labs 继续推动智能上限，open weights 快速吸收突破并降低成本，Applied AI 公司通过 domain evals、workflow context 和模型路由获得差异化，企业自身更重要的工作不是每家公司训练一个模型，而是把 AI 接到正确的数据和流程上。
Aaron also highlighted a model-routing pattern: use frontier intelligence as a manager and cheaper models as workhorses. In one cited case, Fable's stronger delegation reduced total cost because it specified constraints and outcomes, gave feedback, and avoided doing all implementation itself.
Source: https://x.com/levie/status/2076839463410671637
Aaron 还强调了一个模型路由范式：让 frontier intelligence 扮演 manager，让更便宜的模型承担 workhorse 任务。他引用的案例里，Fable 因为更擅长定义约束、目标和反馈，反而降低了总成本，而不是把所有实现都亲自做完。
Ryo Lu, Cursor
Ryo Lu built custom e-reader firmware with Cursor, including Latin and CJK typography, vertical layout, line breaking, large character sets, book/progress sync, and faster rendering with caching. It is a concrete example of AI coding moving beyond web apps into hardware-adjacent personal tools.
Source: https://x.com/ryolu_/status/2076713331113734641
Cursor 的 Ryo Lu 用 Cursor 做了自定义 e-reader firmware，支持 Latin 和 CJK typography、竖排、断行、大字符集、书籍和进度同步，以及带缓存的快速渲染。这是 AI coding 从 web app 扩展到硬件周边个人工具的具体案例。
Zara Zhang
Zara Zhang shared a framework for the three levels of AI adoption in organizations and noted that most companies are at level 2. She also resurfaced a conversation on building in public, audience growth without slop, and vibe coding.
Sources: https://x.com/zarazhangrui/status/2076862290985730481 / https://x.com/zarazhangrui/status/2076860372993388663
Zara Zhang 分享了组织 AI adoption 的三层框架，并判断多数公司目前处在 level 2。她还重新发布了关于 building in public、避免内容 slop 的受众增长，以及 vibe coding 的对谈。
Nikunj Kothari, FPV Ventures
Nikunj Kothari introduced Ramp-Autofill, an open-source Claude Code skill built with Ramp CLI and Fable. It finds receipts from iMessage and Gmail, converts linked receipts to PDFs through Playwright, fills memos from calendar context, learns categorization from prior transactions, verifies its work, and can run as a scheduled job.
Sources: https://x.com/nikunj/status/2076775924650107151 / https://x.com/nikunj/status/2076776777884811671
FPV Ventures 的 Nikunj Kothari 发布了 Ramp-Autofill，一个用 Ramp CLI 和 Fable 构建的开源 Claude Code skill。它可以从 iMessage 和 Gmail 找收据，用 Playwright 把链接里的收据转成 PDF，根据日历上下文补 memo，从历史交易学习分类规则，校验结果，并可作为定时任务运行。
Peter Steinberger
Peter Steinberger said OpenClaw shipped updates across desktop plus iOS and Android, with a Node bump for smooth upgrades. He also shared that moving a maintainer agent to the cloud immediately surfaced multi-agent coordination behavior.
Sources: https://x.com/steipete/status/2076917691139674373 / https://x.com/steipete/status/2076923300593422560
Peter Steinberger 表示 OpenClaw 发布了桌面端以及 iOS、Android 更新，并因为升级流畅性提升了 Node 版本。他还提到把 maintainer agent 搬到云端后，立刻出现了多 agent 协作中的冲突和协调现象。
Sam Altman, OpenAI CEO
Sam Altman said it still "sorta breaks my brain" to see OpenAI models becoming good at design. He also reacted sharply to model access and downgrade policies, signaling that model routing, access tiers, and perceived capability consistency remain live issues in the AI product layer.
Sources: https://x.com/sama/status/2076823209589313910 / https://x.com/sama/status/2076824870072238299
OpenAI CEO Sam Altman 表示，看到 OpenAI 模型终于擅长设计，仍然让他有点 "breaks my brain"。他也尖锐回应了模型访问和降级策略，这说明模型路由、访问层级和用户感知到的能力一致性，仍是 AI 产品层的重要问题。
PODCASTS
The MAD Podcast with Matt Turck: Inside Nemotron & NVIDIA's AI Lab | Bryan Catanzaro
The takeaway: NVIDIA's Bryan Catanzaro argues that open AI is not a side quest to closed frontier labs, but a necessary way for companies to connect intelligence to their own secrets, workflows, and data. He frames progress less as "open versus closed" and more as an ecosystem moving fast together, with open models essential for customization, privacy-sensitive deployment, and domain-specific advantage. One memorable line: "If you accept as the truth that we're gonna be running at the limit, then what that means is that the way to get more intelligence is to be more efficient."
Source: https://www.youtube.com/@DataDrivenNYC/videos
核心 takeaway：NVIDIA 的 Bryan Catanzaro 认为 open AI 不是 closed frontier labs 的旁支，而是企业把智能接入自身秘密、流程和数据的必要路径。他不把问题理解成 "open vs closed"，而是认为整个生态都在高速前进，其中 open models 对定制化、隐私敏感部署和领域优势至关重要。最值得记住的一句话是："If you accept as the truth that we're gonna be running at the limit, then what that means is that the way to get more intelligence is to be more efficient."
He also pushes back against the idea that open progress is simply copied from closed labs or from one country to another. His view is that open communities move faster because people can learn from each other, and China has been especially effective at showing how open model work can accelerate an ecosystem. For builders, the implication is clear: the opportunity is not just choosing the best model, but designing systems that can absorb new open-weight capabilities quickly.
Source: https://www.youtube.com/@DataDrivenNYC/videos
他也反对把 open progress 简化成从 closed labs 或某个国家复制成果。他的观点是：开放社区因为可以互相学习，所以推进速度更快，而中国在展示 open model 如何加速生态方面尤其有效。对 builders 来说，启发很明确：机会不只是选择最强模型，而是设计能快速吸收新 open-weight 能力的系统。
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders