AI Builders
Digest

OpenAI Codex & ChatGPT 负责人 Thibault Sottiaux 表示这是他在 OpenAI 最好的一天之一，充满乐趣与强度，未来一片光明。他特别点赞 Sergio 正在 ChatGPT 中打造的协作空间功能，每天都在飞速进步。

OpenAI Codex & ChatGPT lead Thibault Sottiaux called it one of his best days at OpenAI yet, full of fun and intensity with a bright future ahead. He highlighted Sergio cooking up collaborative space right inside ChatGPT, improving leaps and bounds every day.

Meta AI 高级总监 Madhu Guru 指出，团队最常见的错误是把评估当成 Agent 建好后的额外 QA 步骤。AI 产品本质不同，你的评估就是你的产品规格。Box CEO Aaron Levie 补充，Agent 落地瓶颈在于在真实工作环境中测试、调优和优化，包括文件、CRM、邮件等。每个企业最终都需要有人专门管理评估和模拟环境基础设施，这是巨大空间。

Meta Sr Director of AI Madhu Guru warned that the most common mistake is treating evals as an extra QA step after the agent is built. AI products are fundamentally different: your evals are your product spec. Box CEO Aaron Levie added that agent adoption is bottlenecked by the ability to test, tune and optimize on real work environments (files, CRM, email). Every enterprise will need people managing evals and simulated environments, a huge space.

Replit CEO Amjad Masad 指出美国正在开源权重模型领域迎头赶上。他同时为大量家族企业使用 Replit 供电感到自豪。

Replit CEO Amjad Masad noted that the US is catching up on open-weights models. He also expressed pride in how many family-run businesses are powered by Replit.

Vercel CEO Guillermo Rauch 发布了 gdp-ts，这是一个库、linter 和 AI skill，用于更安全的 API 设计。敏感函数要求调用者提供已完成授权检查的“证明”，类型检查器在编译时验证这些证明，防止团队和 Agent 发布灾难性安全漏洞。随着 Agent 写代码量超过人类审查能力，这种硬约束反而成为优势，Agent 在紧约束循环中表现更好。README 以真实 Vercel API 约束为例：修改 Project 密码需要特定角色和权益证明。

Vercel CEO Guillermo Rauch introduced gdp-ts, a library, linter and AI skill for safer API design. Sensitive functions require 'proofs' that the caller performed an authorization check; the typechecker verifies them at compile time, stopping teams and agents from shipping catastrophic security bugs. Agents now write more code than humans can review and thrive under hard constraints that frustrate people. The README models a real Vercel constraint: changing a Project password needs proof of a certain role plus entitlement.

Anthropic Claude Code 的 Thariq 分享了他最喜欢的模式名称 “local hands”：Claude 在云端运行却能访问本地文件，该能力也即将进入 Cowork。他指出这种规划方式比原始 HTML 更省 token，模型无需重复生成状态机、图表、代码片段等常见组件。

Anthropic Claude Code’s Thariq shared his favorite name for the pattern “local hands”: Claude runs in the cloud but can access your files locally, also coming to Cowork. A big advantage of this planning style is far higher token efficiency than raw HTML, because the model does not need to remake components or logic for common things like state machines, diagrams and code snippets.

实用 AI 教程作者 Peter Yang 分享了他用 Gemini Live API 构建的实时语音日语学习 App 完整教程，包括用 spec skill 做关键设计、设置 Gemini Live 语音 API，以及用 Nano Banana 生成立体模型艺术。用户可跟着做，在旅行前学会任意语言的 100 个短语。

Practical AI tutorials creator Peter Yang released a full tutorial on the AI language-learning app he built that teaches Japanese via live voice calls. It covers using his spec skill for key designs, setting up Gemini Live APIs for voice, and creating diorama art with Nano Banana. Followers can build their own version to learn 100 phrases for any language before a trip.

曾为 Cursor、Notion、Stripe 设计的 Ryo Lu 发布了 ryOS Subtitles Chrome 扩展，让语言学习者可在 Netflix 上同时显示任意两种语言的字幕，并为日语、中文、韩语提供发音指南，支持自定义样式。

Designer Ryo Lu (ex-Cursor, Notion, Stripe) released ryOS Subtitles for Chrome. It lets language learners watch Netflix with subtitles in any two languages plus pronunciation guides for Japanese, Chinese and Korean, with customizable styles and hand-picked defaults.

Y Combinator 总裁兼 CEO Garry Tan 指出，实验室的 harness 有激励去烧 token，而创业公司的 harness 则具备真实效用，例如 Grep 能自动把 token 消耗换成可重复的确定性测试代码。他还提到 AGI Science Loops 即将到来，Halmos 正在为这些循环构建 Muse/Instinct。

Y Combinator President & CEO Garry Tan observed that harnesses from labs have an incentive to burn tokens, while harnesses from startups deliver real utility, such as Grep automatically swapping token burn into deterministic tested code that is repeatable by observing agent use. He also noted that AGI Science Loops are coming and Halmos is building the Muse/Instinct for them.

设计师 Ryo Lu 写下长文纪念 Steve Jobs。他从小玩米色电脑，后来在加拿大看到 iMac G4，被其美感震撼，于是把 PC 改造成 Mac 风格并由此进入界面设计。他反复观看 Jobs 的 keynote DVD，第一台 iPod 是黑色 Nano。Jobs 去世后他感到行业陷入不断打磨与供应链优化的循环，缺少真正用心的作品。他仍常问自己 “Steve 会怎么做”，并引用那句 “那些疯狂到以为自己能改变世界的人，正是真正改变世界的人”。

Designer Ryo Lu wrote a heartfelt tribute to Steve Jobs. Growing up with beige boxes, he was stunned by the iMac G4 in Canada and taught himself interface design by making his PC look like a Mac. He burned DVDs of Jobs’ keynotes and studied them obsessively; his first iPod was a black Nano. Since Jobs’ death he has felt the industry stuck in constant refinement and supply-chain optimization, with little room for something truly great and heartfelt. He still asks “what would Steve do?” and keeps the line “the people who are crazy enough to think they can change the world, are the ones who do.”

SPC 普通合伙人 Aditya Agarwal 认为，能够在本地运行 Muse/Dot “电脑”会很有意思，感觉那会是比日益锁定的 macOS 更好的 Agent 环境。

SPC General Partner Aditya Agarwal suggested it would be interesting to run the Muse/Dot “computer” locally, feeling that would be a better agent environment than increasingly locked-down macOS.