AI Builders Digest - 2026-07-08

Stats: xBuilders=16, totalTweets=34, podcastEpisodes=1

X / TWITTER

Swyx, AI Engineer / Latent Space

Swyx highlighted Anthropic's J-space paper as important because it moves from correlation to intervention: the researchers can change a model's reasoning midstream, and the model can detect that the intervention happened. His read is that this sits close to eval awareness and makes mechanistic control of reasoning more concrete.

Swyx 认为 Anthropic 的 J-space paper 最值得注意的地方，是它从相关性走向了干预：研究者可以在推理中途改变模型的主题，而模型还能识别自己被干预过。他的判断是，这已经接近 eval awareness，也让“控制模型推理”从概念变得更具体。

Link: https://x.com/swyx/status/2074344727202463832

Boris Cherny, Claude Code at Anthropic

Boris Cherny pointed to Anthropic's first public telling of how Claude Code was built and launched, tracing the product back to safety research. The notable builder signal is the posture: even after the launch and adoption, the team frames Claude Code as “1% done,” meaning the current coding-agent wave is still early infrastructure, not a finished category.

Boris Cherny 转发了 Anthropic 第一次系统讲述 Claude Code 从构建到发布的故事，并把它的源头追溯到安全研究。真正值得关注的是团队姿态：即便已经发布并获得使用，他们仍然说 Claude Code 只是“完成了 1%”，这说明 coding agent 还处在早期基础设施阶段，而不是成熟品类。

Link: https://x.com/bcherny/status/2074247226038063316

Thibault Sottiaux, Codex and ChatGPT at OpenAI

Thibault Sottiaux noted that teams building ChatGPT, Codex, and OpenClaw would be present at an upcoming event, with “a few surprises” planned. This is mostly an ecosystem signal: OpenAI is putting coding agents, ChatGPT product work, and OpenClaw-style agent tooling in the same builder conversation.

Thibault Sottiaux 提到，构建 ChatGPT、Codex 和 OpenClaw 的团队会出现在即将举行的活动上，并准备了一些“surprises”。这更像是生态信号：OpenAI 正在把 coding agent、ChatGPT 产品能力和 OpenClaw 式 agent 工具放到同一个 builder 语境里。

Links: https://x.com/thsottiaux/status/2074195169990357398, https://x.com/thsottiaux/status/2074209421799166138

Peter Yang, AI educator and product writer

Peter Yang shared concrete prompts for using Fable 5 before it leaves Claude subscriptions, including a prompt for finding “Fable-worthy work” across projects, docs, and memory. The useful pattern is not model worship, but task triage: reserve frontier models for work where context, judgment, and leverage justify the cost.

Peter Yang 分享了在 Fable 5 离开 Claude 订阅前值得尝试的具体 prompts，其中包括让模型浏览项目、文档和 memory，找出真正值得用 Fable 的任务。这里的重点不是追新模型，而是任务分层：把 frontier model 留给那些需要上下文、判断力和高杠杆的工作。

Link: https://x.com/petergyang/status/2074206798631071796

Nan Yu, Head of Product at Linear

Nan Yu argued that brute-force work hours used to be a stronger advantage because so much time went into tedious programming tasks. With agents, even early-stage startups can now run more work in parallel, so the old 996-style intensity still matters in places, but is less universally effective than before.

Linear 产品负责人 Nan Yu 的判断是，过去长时间工作之所以有效，很大一部分是因为大量时间花在繁琐编程任务上。现在有了 agents，即使早期创业公司也能并行推进更多工作，所以 996 式强度仍有适用场景，但不再像以前那样普遍有效。

Links: https://x.com/thenanyu/status/2074133468007587932, https://x.com/thenanyu/status/2074258147015897357

Cat Wu and Thariq, Claude Code team at Anthropic

Cat Wu and Thariq both amplified Anthropic's Claude Code retrospective. Thariq also clarified a concrete deadline around Claude availability, but the durable signal is that Anthropic is turning Claude Code's origin story into public product narrative: safety research, early team practice, and user feedback are being packaged as the product's legitimacy.

Cat Wu 和 Thariq 都转发了 Anthropic 关于 Claude Code 的复盘文章。Thariq 还补充了 Claude 可用性的具体截止时间，但更长期的信号是：Anthropic 正在把 Claude Code 的起源故事塑造成公开产品叙事，把 safety research、早期团队实践和用户反馈包装成产品可信度的一部分。

Links: https://x.com/_catwu/status/2074258446686536167, https://x.com/trq212/status/2074186977147273540, https://x.com/trq212/status/2074185669598237047

Amjad Masad, CEO of Replit

Amjad Masad shared two strong Replit signals: a real estate company reportedly saved $100k by replacing Salesforce with a Replit-built CRM, and Replit's agent is improving rapidly because the team “closed the loop” into self-improvement. Together, these point to a practical wedge for AI software creation: internal tools and vertical workflows where replacement value is obvious and iteration loops are tight.

Replit CEO Amjad Masad 分享了两个强信号：一家亚特兰大的房地产公司用 Replit 构建的 CRM 替代 Salesforce，据称省下了 10 万美元；同时 Replit agent 之所以进步很快，是因为团队把反馈闭环做进了 self-improvement。两者合起来说明，AI 软件生成最现实的切入口仍然是内部工具和垂直流程：替代价值明确，迭代闭环足够短。

Links: https://x.com/amasad/status/2074274666709987663, https://x.com/amasad/status/2074257906594177279, https://x.com/amasad/status/2074353874996211831

Guillermo Rauch, CEO of Vercel

Guillermo Rauch framed the ultimate test for coding AI as whether software as a whole gets better: faster shipping, new apps and games, fewer bugs, and users expanding usage. He also pointed to eve shipping evals out of the box, arguing that unlike web frameworks where testing became an ecosystem choice, agents need evals as a built-in primitive.

Vercel CEO Guillermo Rauch 认为，coding AI 的终极测试不是 token 用量，而是整个软件行业是否真的变好：公司是否更快 shipping，是否出现过去做不出来的 app 和 game，软件 bug 是否减少，用户是否持续扩大使用。他还提到 eve 默认内置 evals，因为和 Web framework 不同，agent 的 evals 不是生态可选项，而是基础能力。

Links: https://x.com/rauchg/status/2074222247548735996, https://x.com/rauchg/status/2074287795028512773

Aaron Levie, CEO of Box

Aaron Levie laid out a pragmatic enterprise AI stack: frontier models stay important for new use cases, orchestration, and complex planning, while mature predictable workflows can shift some tokens to cheaper open or task-specific models. The applied AI layer becomes the control plane that evaluates workflows, chooses model mixtures, and eventually creates domain-specific training loops.

Box CEO Aaron Levie 给出了一个务实的企业 AI 分层：frontier model 仍然负责新用例、编排和复杂规划；一旦工作流成熟且可预测，就可以把一部分 token 迁移到更便宜的 open model 或任务专用模型。真正关键的是 applied AI layer，它负责评估工作流、选择模型组合，并逐步形成领域训练闭环。

Link: https://x.com/levie/status/2074163686990913576

Zara Zhang, AI builder

Zara Zhang recommended binge-watching recent talks from AI Engineer, Cursor Compile, and Figma Config, arguing that high-quality conference knowledge is under-consumed relative to its value. Her sharper point is about role collapse: in AI-native work, everyone increasingly needs to think like an engineer, PM, and designer.

Zara Zhang 推荐系统观看 AI Engineer、Cursor Compile 和 Figma Config 最近的公开视频，认为这些高质量 conference talks 的价值被低估了。更关键的判断是角色边界正在坍缩：在 AI-native 工作方式里，每个人都越来越需要同时像 engineer、PM 和 designer 一样思考。

Links: https://x.com/zarazhangrui/status/2074304295097561490, https://x.com/zarazhangrui/status/2074305070955639077

Nikunj Kothari, Partner at FPV Ventures

Nikunj Kothari pushed back on over-scientific VC sourcing narratives: venture investing is about outliers, but many firms still try to prove differentiation through proprietary algorithms and abstract metrics. His practical model is expanding the surface area for luck while building a prepared mind around a few themes 6 to 12 months early.

FPV Ventures 合伙人 Nikunj Kothari 反思了 VC 对“科学化 sourcing”的迷恋：风险投资本质上投的是 outliers，但很多机构仍试图用 proprietary algorithm 和各种指标证明自己不同。他更实际的方法是扩大幸运发生的接触面，同时围绕少数提前 6 到 12 个月的主题建立 prepared mind。

Link: https://x.com/nikunj/status/2074141483356340475

Peter Steinberger, OpenClaw and OpenAI

Peter Steinberger raised a concrete operating question for the new engineering world: how should teams run AI-assisted engineering interviews? That is a useful signal because hiring loops are now lagging behind the work itself; if daily engineering is agent-assisted, interview design has to test judgment, delegation, debugging, and review instead of pretending AI is absent.

Peter Steinberger 提出了一个很现实的工程管理问题：现在应该如何进行 AI-assisted engineering interview？这值得关注，因为招聘流程已经落后于真实工作方式；如果日常工程已经由 agent 辅助，面试就应该测试判断、委托、调试和 review 能力，而不是假装 AI 不存在。

Links: https://x.com/steipete/status/2074380549318443311, https://x.com/steipete/status/2074210475777364197

Claude, Anthropic

Claude shared a short history of how Claude Code came to be, told by the builders and early users behind it. The positioning is clear: Claude Code is not just another IDE assistant, but a product lineage Anthropic wants tied to research, product craft, and early adopter practice.

Claude 官方发布了 Claude Code 的简短历史，由构建者和早期用户讲述。它的定位很清楚：Claude Code 不只是又一个 IDE assistant，而是 Anthropic 希望与研究、产品手感和早期用户实践绑定在一起的产品谱系。

Link: https://x.com/claudeai/status/2074244664199115201

PODCASTS

AI & I by Every: Building a School Where AI Models Learn About Humanity

The takeaway: Edwin Chen frames Surge as a “school for AGI,” where models learn taste, judgment, and expert reasoning, not just labeled answers. The interesting tension is existential and practical at once: if scaling continues, models may become capable of much of what humans do, but the right goal for AI products should not be engagement. Chen argues that useful AI should sometimes push back, end loops, and help humans become more capable rather than keep them chatting forever.

Every 的 AI & I 这一期里，Surge 创始人 Edwin Chen 把 Surge 描述成一所“AGI 学校”：模型在这里学习的不只是标注答案，而是 taste、judgment 和专家级推理。真正有价值的张力同时是存在论和产品论：如果 scaling 持续推进，模型会越来越能做人类能做的事，但 AI 产品的目标不应该是 engagement。Chen 认为，好的 AI 有时应该 push back，结束无效循环，帮助人变得更强，而不是让人无限聊天。

Link: https://www.youtube.com/watch?v=omX6wrLuX08

Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders