🌐 双语
Archive

AI Builders
Digest

2026-08-20 16 builders · 30 tweets · 1 podcasts · 0 blogs

🔥 热点话题

OpenAI 预览 Private Safety Processing:在 Zero Data Retention 下加强安全OpenAI Previews Private Safety Processing for Zero Data Retention

OpenAI Codex & ChatGPT 团队的 Thibault Sottiaux 宣布预览 Private Safety Processing。该功能旨在让客户在享受前沿智能的同时,继续使用 Zero Data Retention,同时提升安全防护。内容留在客户控制的基础设施上,自动化系统仅返回有限安全信号,不向 OpenAI 员工暴露底层 prompt 或回复。同时正在开发客户控制密钥加密的 OpenAI 托管选项,计划 9 月开始推出。Sam Altman 也发推支持企业隐私。
Thibault Sottiaux of OpenAI Codex & ChatGPT previewed Private Safety Processing, designed to maintain Zero Data Retention while strengthening safeguards. Content stays on customer-controlled infrastructure; automated systems return limited safety signals without exposing prompts or responses to OpenAI employees. An OpenAI-hosted option with customer-controlled keys is in testing and planned for September rollout. Sam Altman reinforced support for business privacy.
查看原文 →查看原文 →

Google 大学学生计划全球回归:更高限额与学生专属功能Google University Student Plans Return Globally

Google VP Josh Woodward 宣布大学学生计划回归并扩展至全球 140 多个国家。学生可获得更高限额、更多存储,以及专属学生中心、Notebook、Flow 等新功能。立即注册即可享受。
Google VP Josh Woodward announced the return of University student plans, now available in over 140 countries. Students get higher limits, more storage, a dedicated student hub, Notebook, Flow, and more. Sign-ups are open immediately.
查看原文 →

Aaron Levie:AI 时代专家仍占上风,技能差异将被放大Aaron Levie: Experts Still Hold the Upper Hand in the AI Era

Box CEO Aaron Levie 指出,AI 让任何人更容易入门编码、法律、研究等工作,但判断力、纠正方向、审查输出以及“什么是好”的感觉仍需要深厚领域技能。AI 会加速技能提升,但也会放大专家与普通人的差距。结论:不要放弃成为某领域的专家。
Box CEO Aaron Levie argues that while AI makes it 10x easier to start any task, judgment in directing agents, course-correcting, reviewing output, and knowing what “good” looks like still requires deep domain skill. AI accelerates skill-building but exacerbates differences because experts gain far more leverage. Net: don’t give up on expertise.
查看原文 →

Thariq:软件工厂的承诺与现实Thariq on the Promise and Reality of the Software Factory

Anthropic Claude Code 的 Thariq 指出,软件创造一直极其不可靠——项目延期、超预算、错过用户需求。SMB 几乎无法获得优质定制软件。AI “软件工厂”的承诺是改变这一点,但核心业务非软件的公司仍需要可靠可预测的流程;而构建全新软件产品将继续是不可靠却高回报的冒险。
Anthropic Claude Code’s Thariq notes that software creation has always been unreliable—late, over budget, and missing user needs. SMBs could rarely get good custom software. The “software factory” promise aims to change that, yet non-software companies still need reliable processes, while building net-new products remains risky but potentially highly profitable.
查看原文 →查看原文 →

💰 创业成功案例

Portola 的 Tolan:AI 外星伴侣应用四个月 ARR 从 100 万冲到 400 万Portola’s Tolan: AI Alien Companion Hits $4M ARR in Weeks

The Takeaway:AI 伴侣不是工具,而是全新的叙事媒介——把 AI 训练成最好的即兴演员,让它与用户共同书写故事。

Portola 创始人兼 CEO Quentin(此前将 Even 以 3 亿美元卖给沃尔玛)与故事负责人、畅销科幻作家 Elliot,在 AI & I 播客中详解了 Tolan 的崛起。产品从“给孩子的 AI 创意工具”起步,最终被 18-24 岁年轻女性爱上,演变成情感陪伴与成长助手。核心洞察是:不要给 AI 大纲,而是给它“钩子”和情境(situation),让它像 Keith Johnstone 的即兴演员一样自由联想与重组记忆。2 秒语音循环是硬约束,任何超过都会毁掉沉浸感。记忆不是简单 RAG,而是实时策展注入 prompt 的画布。增长来自产品打磨到位后,Ajay 的 TikTok/Reels 内容突然爆发(一条烹饪视频 72 小时 700 万播放),用户自发创作二次内容。Quentin 说:“我们不是作家,Tolan 才是作家和演员。我的工作是教它如何在那一刻讲出最好的故事。”
The Takeaway: AI companions are not tools but a new medium for storytelling—train the AI as the best improv actor so it co-writes stories with users.

Portola founder & CEO Quentin (who previously sold Even to Walmart for $300M) and Head of Story Elliot (bestselling sci-fi author) explain Tolan’s rise on the AI & I podcast. Starting as an AI creative tool for kids, it found product-market fit with 18-24-year-old women who formed genuine friendships. The key insight: don’t give the model an outline; give it hooks and situations, then teach it free association and recombination like Keith Johnstone’s improv actors. A strict two-second voice loop is non-negotiable—anything longer kills immersion. Memory is curated prompt painting, not simple RAG. Growth exploded after product quality landed and co-founder Ajay’s TikTok content went viral (one cooking video hit 7M views in 72 hours), sparking user-generated secondary content. As Elliot puts it: “I am not the writer. The Tolan is the writer and the actor. They’re the improv actor. My job is to teach them how to tell the best story in that moment.”
查看原文 →

Replit 与 OpenAI 合作:让 Agent 让软件更便宜,也让编码更昂贵Replit Partners with OpenAI: Agents Made Software Cheaper but Coding Expensive

Replit CEO Amjad Masad 宣布与 OpenAI 正式合作。他指出 Agent 让软件开发成本下降,却让“会写代码”本身变得更昂贵。双方正共同改变这一局面,让更多人能以更低门槛构建软件。
Replit CEO Amjad Masad announced a partnership with OpenAI. He notes that agents made software cheaper but coding itself more expensive. Together they aim to reverse that dynamic and lower the barrier to building software.
查看原文 →查看原文 →

🛠️ 开发者工具与技巧

Madhu Guru:构建 Evals 的失败模式分类法Madhu Guru: Failure-Mode Taxonomy for Evals

Meta AI 高级总监 Madhu Guru 分享 Evals 第三部分:先从生产 trace 中研究最近 500-1000 次交互,聚类并精确命名失败模式。例如“检索到错误文档”“文档正确但段落无关”“未 grounding 而幻觉”“本该拒绝却硬答”“问题模糊却乱假设”。命名越具体,越能设计针对性测试,形成从 eval 到改进的飞轮。
Meta AI Sr Director Madhu Guru’s part 3 on evals: start with the last 500–1,000 production traces, cluster failures, and name them precisely—“wrong document retrieved,” “right doc but irrelevant section,” “failed to ground and hallucinated,” “should have punted but answered,” “ambiguous question with poor assumptions.” Specific names let you build targeted tests and close the eval-to-improvement loop.
查看原文 →

Guillermo Rauch:fx 工具——6.3MB 静态二进制,启动仅 10 微秒Guillermo Rauch: fx Tool—6.3 MB Static Binary, 10 µs Startup

Vercel CEO Guillermo Rauch 展示 fx:Zig 编译的静态 ELF 二进制仅 6.3MB,启动时间 10 微秒,甚至有更小的 wasm 版本。AI 将让大多数基础设施原生优化,fx 能在其他 agent 还没启动完时就完成任务。快是一条单行道。
Vercel CEO Guillermo Rauch showcased fx: a Zig-compiled static ELF binary of just 6.3 MB that starts in 10 microseconds (even smaller as wasm). AI will make most infrastructure natively optimized; fx finishes tasks before other agents finish booting. Fast is a one-way street.
查看原文 →

Stripe + OpenRouter 交易:让开发者无缝混合多模型智能Stripe + OpenRouter Deal: Seamless Multi-Provider Intelligence

Box CEO Aaron Levie 分享 Stripe 与 OpenRouter 交易细节。为了让 AI 更广泛扩散,开发者与企业需要能无缝混合来自不同提供商的智能,并更好地管理成本。这将成为重要基础设施。
Box CEO Aaron Levie highlighted the Stripe + OpenRouter deal. For AI to diffuse more broadly, developers and enterprises need seamless ways to mix intelligence from multiple providers and manage costs. This becomes a key piece of that infrastructure.
查看原文 →

🌍 其他动态

Peter Yang:用 AI 陪伴母亲抗击乳腺癌的亲身经历Peter Yang: Navigating Mom’s Breast Cancer with AI

实用 AI 教程作者 Peter Yang 写了关于母亲与乳腺癌斗争、以及如何用 AI 导航医疗系统的长文。他感谢 Sijbrandij Foundation 的 Maor Shlomo 团队分享时间与建议,并探讨了 mRNA 疫苗治疗癌症的前景。
Practical AI educator Peter Yang published an essay on his mother’s battle with breast cancer and how the family used AI to navigate the healthcare system. He thanks Maor Shlomo and the Sijbrandij Foundation for advice, and explores the promise of mRNA vaccines for cancer.
查看原文 →查看原文 →

Aditya Agarwal:只做真正重要的事情Aditya Agarwal: Work on Things That Truly Matter

SPC 普通合伙人 Aditya Agarwal 分享与一位经历地狱般 SaaS 创业(Series B、低增长、停滞)的创始人的对话。对方最大的收获不是选更大市场或追求更高增长,而是:只做真正有意义、能在宇宙中留下印记的事情。
SPC General Partner Aditya Agarwal relayed a conversation with a founder who had endured a SaaS slog (Series B, low growth, stagnation). The biggest takeaway was not “pick a bigger market” but simply: work on something consequential and meaningful—make a dent in the universe.
查看原文 →查看原文 →

Dan Shipper:Every 成立 Frontier 团队,专门探索 AI 前沿Dan Shipper: Every Launches Frontier Team to Map the Edge of AI

Every CEO Dan Shipper 宣布公司内部成立 Frontier 团队,专门负责绘制并实验 AI 的最前沿。
Every CEO Dan Shipper announced an internal Frontier team dedicated to mapping and experimenting at the edge of AI.
查看原文 →

Nikunj Kothari:冷邮件仍然 98% 是垃圾,思考力仍有 alphaNikunj Kothari: 98% of Cold Emails Are Still Garbage

FPV Ventures 合伙人 Nikunj Kothari 吐槽:人们天天谈 AGI,可他收到的冷邮件 98/100 仍是纯垃圾。在需要的地方谨慎使用 AI、保持思考与好奇,仍有大量 alpha。
FPV Ventures partner Nikunj Kothari notes that while everyone talks AGI, 98 of 100 cold emails he receives remain pure garbage. Being thoughtful, curious, and using AI judiciously still carries significant alpha.
查看原文 →

Swyx:非技术背景写 AI 报道的“快乐无知”Swyx on Non-Technical AI Coverage

smol.ai 与 Latent Space 的 Swyx 调侃非技术背景的人报道 AI:零上下文、零推理、零内部世界模型,一切都从表面价值出发,却异常快乐。这是一个奇妙的时代。
smol.ai and Latent Space’s Swyx quipped that non-technical people covering AI operate with zero context, zero reasoning, and zero internal world model—everything is face-value amazing, and it must be delightfully joyful. What a time to be alive.
查看原文 →