AI Builders Digest — 2026-11-03

X / TWITTER

Andrej Karpathy

Andrej Karpathy built HackerTube in roughly two hours with Claude Code and Opus 5.5, filtering 43 million Hacker News comments into 6,454 repeatedly recommended videos. The striking part is not the recommendation system itself but the workflow: he handled only the unavoidable browser steps while Claude built, secured, deployed, and optimized the service, turning a project that once might have taken weeks into a breakfast-sized experiment.

Andrej Karpathy 用 Claude Code 和 Opus 5.5 在大约两小时内做出了 HackerTube，把 Hacker News 二十年来的 4,300 万条评论筛成 6,454 个高频推荐视频。真正值得关注的不是推荐系统本身，而是工作方式：他只处理必须由人在浏览器中完成的环节，其余构建、加固、部署和优化均交给 Claude，把过去需要数周的项目压缩成一次早餐时间的实验。

https://x.com/karpathy/status/2116957135661945283

Peter Yang

AI educator Peter Yang argues that different tools still dominate different jobs: ChatGPT plus Codex for browser and computer use, Claude for design, video editing, and taste, Grok Bot for persistent agents and scheduled work, with Gemini underrated for writing. He also highlights a product-design split: humans should remain in the loop for creating, learning, and consuming, while many utilitarian interfaces may be better operated directly by agents.

AI 教育者 Peter Yang 认为，目前不同工具仍各有明确优势：ChatGPT 加 Codex 擅长 browser 和 computer use，Claude 擅长设计、视频编辑与“品味”，Grok Bot 适合常驻 agent 和定时任务，而 Gemini 的写作能力被低估。他还指出产品设计正在分化：创造、学习和内容消费仍需要人直接参与，但许多工具型界面更适合让 agent 直接操作。

https://x.com/petergyang/status/2116946320674152810

https://x.com/petergyang/status/2116987415319740756

Thariq

Former Claude Code builder Thariq is prototyping a game with AI-generated sprites, replacing 3D models and adding a fire-wielding taekwondo character. He expects a vertical slice within a week, a compact example of generative assets shortening the iteration loop for a solo game builder.

前 Claude Code 成员 Thariq 正在用 AI-generated sprites 制作游戏原型，用生成式精灵图替代 3D 模型，并加入一名使用火焰攻击的跆拳道角色。他计划一周内完成 vertical slice，展示了生成式素材如何明显缩短独立游戏开发者的迭代周期。

https://x.com/trq212/status/2116976789046350040

https://x.com/trq212/status/2116977363728867664

Replit CEO Amjad Masad

Replit CEO Amjad Masad warns that heavily subsidized AI subscriptions are unlikely to last, implying users should prepare for repricing as providers confront real inference costs. More concretely, Replit now lets teams turn feedback-widget submissions directly into prompts for agents to build from, tightening the loop from user request to shipped product change.

Replit CEO Amjad Masad 警告，依靠高额补贴维持的 AI 订阅难以长期持续，用户应为供应商面对真实 inference 成本后的重新定价做好准备。更具体的产品进展是，Replit 现在可以把 feedback widget 收到的用户意见直接变成 agent 的构建 prompt，进一步压缩从用户需求到产品更新的闭环。

https://x.com/amasad/status/2117010858845089805

https://x.com/amasad/status/2116998155925360965

Vercel CEO Guillermo Rauch

Vercel CEO Guillermo Rauch makes the strongest sovereignty argument of the day: own your harness, data, connections, runtime, and agents, even if frontier-scale models cannot yet run locally. His practical prescription is to keep model options open, support local-model proliferation, and embrace open source; meanwhile, his tests suggest Gemini is already performing well on two high-complexity projects.

Vercel CEO Guillermo Rauch 提出了当天最鲜明的“智能主权”主张：即便 frontier 级大模型暂时无法在本地运行，也要掌握自己的 harness、数据、连接、runtime 和 agents。他的实际建议是保持模型选择权、推动 local models 普及并拥抱 open source；与此同时，他在两个高复杂度项目中的测试显示 Gemini 表现出色。

https://x.com/rauchg/status/2117000819623243925

https://x.com/rauchg/status/2116988333096370443

Box CEO Aaron Levie

Box CEO Aaron Levie predicts that many knowledge workers will shift from individual contributors using AI for incremental leverage to managers of parallel agent teams. In information-heavy fields such as software, cybersecurity, law, biotech, and engineering, output can become decoupled from hours worked; organizations that redesign operations around this early may compound their advantage.

Box CEO Aaron Levie 预测，大量知识工作者将从“用 AI 获得增量效率的个人贡献者”转变为“管理并行 agent 团队的人”。在软件、网络安全、法律、生物科技和工程等信息密集型领域，产出将逐渐与工时脱钩；越早围绕这种模式重构运营的组织，越可能形成复利优势。

https://x.com/levie/status/2117134917847093491

Y Combinator President and CEO Garry Tan

Y Combinator President and CEO Garry Tan highlights Particle-1, claiming agentic retrieval performance comparable to much larger models at one twenty-third of the price. He also points to practical robotics and hard tech as increasingly immediate opportunities, reinforcing the view that AI startup momentum is moving beyond software-only products.

Y Combinator 总裁兼 CEO Garry Tan 重点推荐了 Particle-1，称其 agentic retrieval 能力可媲美更大的模型，但价格只有后者的二十三分之一。他同时强调 practical robotics 与 hard tech 正迅速从远期叙事变成现实机会，说明 AI 创业动能正在从纯软件产品向实体技术扩展。

https://x.com/garrytan/status/2117137390016073977

https://x.com/garrytan/status/2116941191526895975

https://x.com/garrytan/status/2116909591607099793

PODCASTS

The feed reported one new Unsupervised Learning episode featuring Modal founder and CEO Eric Seufert, but supplied only a channel URL rather than a specific episode URL. It is omitted to preserve the digest's source-link integrity rule.

Feed 报告了一期由 Modal 创始人兼 CEO Eric Seufert 参与的 Unsupervised Learning 新节目，但只提供了频道地址，没有具体单集链接。为遵守原始来源链接完整性规则，本期暂不收录。

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