AI Builders Digest — 2026-10-25

2026-10-25

AI Builders Digest — 2026-10-25

X / TWITTER

OpenAI Codex & ChatGPT engineer Thibault Sottiaux

Thibault Sottiaux says Codex CLI usage has doubled since DevDay. Recent updates add reusable prompt presets, inline model and permission controls, enterprise MCP sign-in, more reliable long-session resume, and corporate-proxy authentication on Windows. Codex and ChatGPT also now preserve cache when reasoning settings change mid-conversation, while Dots calls connect faster.

OpenAI Codex 与 ChatGPT 工程师 Thibault Sottiaux 表示，Codex CLI 自 DevDay 以来使用量翻倍。近期更新加入可复用的 prompt presets、内联 model 与 permission 控制、企业级 MCP 登录、长会话恢复稳定性改进，以及 Windows 企业代理认证；Codex 和 ChatGPT 在对话中途切换 reasoning 设置时也不再使缓存失效，Dots 通话连接速度得到提升。

Claude Code engineer Thariq

Thariq announced that he has left Anthropic after more than a year helping Claude Code scale beyond his expectations. He framed Anthropic's mission as a "race to the top," where responsible competition raises the standard across the industry, and said he will take a break before starting something new.

Claude Code 工程师 Thariq 宣布离开 Anthropic。他在一年多时间里参与推动 Claude Code 实现超预期增长，并将 Anthropic 的使命概括为一场“向上竞争”：通过负责任的竞争提高整个行业的标准。他将先休息一段时间，再投入新的项目。

https://x.com/trq212/status/2113844051938054167

Replit CEO Amjad Masad

Replit is bringing Jeff and the Chroma team into the company to tackle one of the hardest agent problems: supplying enough context and durable memory for agents to do real work. Masad calls this a major step toward the "self-driving company." Replit also launched Replit for Government and hired former AWS public-sector leader Michael Greenwald to lead the effort.

Replit 将 Jeff 及 Chroma 团队纳入公司，重点解决 agent 最困难的问题之一：为其提供完成真实工作所需的上下文与持久记忆。Amjad Masad 将其称为迈向“自动驾驶公司”的重要一步。Replit 同时发布 Replit for Government，并引入前 AWS 公共部门负责人 Michael Greenwald 推动政府业务。

Vercel CEO Guillermo Rauch

Guillermo Rauch highlights a long-known but newly practical risk: agents can exploit timing differences in shared-host CPU caches to infer activity from neighboring workloads. Fully eliminating this "noisy neighbor" class of attack would require dedicated hardware per agent, with major economic costs. Vercel Sandbox is instead combining all practical mitigations with continuous monitoring and red-teaming.

Vercel CEO Guillermo Rauch 指出一种已知多年、但如今因 agent 而变得更现实的风险：agent 可以利用共享主机 CPU cache 的时序差异，推断相邻工作负载的活动。若要彻底消除这类“吵闹邻居”攻击，每个 agent 都需要独占硬件，经济成本极高。Vercel Sandbox 当前采用所有可行缓解措施，并结合持续监控与 red-teaming。

https://x.com/rauchg/status/2113759533801877920

Box CEO Aaron Levie

Aaron Levie predicts that agent swarms for cyber defense will become one of AI's largest categories. Security rewards parallel compute directly: swarms can inspect code and infrastructure, test far more risk permutations than human teams, and process volumes of evidence that organizations rarely review quickly enough.

Box CEO Aaron Levie 预测，用于网络防御的 agent swarm 将成为 AI 最大的赛道之一。安全工作能直接从并行算力中获益：agent swarm 可以审查代码与基础设施、测试远超人类团队承载能力的风险组合，并处理组织过去难以及时审阅的海量证据。

https://x.com/levie/status/2113738768645808277

Every CEO Dan Shipper

Dan Shipper argues that the strongest AI teams need only two archetypes: "pirates" who move aggressively and improvise, and "architects" who impose coherent structure. The point is less about headcount than balancing speed with systems thinking.

Every CEO Dan Shipper 认为，最强的 AI 团队只需要两类角色：“海盗”负责激进推进与临场创造，“架构师”负责建立连贯结构。关键不在团队人数，而在速度与系统思维之间形成平衡。

https://x.com/danshipper/status/2113693510629281895

SPC General Partner Aditya Agarwal

Aditya Agarwal points to a recurring pattern in AI progress: major advances often come from solving many problems jointly rather than training separate specialist systems. He says Physical Intelligence cofounder Sergey Levine sees the same principle in robotics: make the problem bigger so a general system can learn across tasks.

SPC General Partner Aditya Agarwal 指出 AI 进步中的一个反复出现的规律：重大突破往往来自联合解决多个问题，而不是为每个任务训练独立的专用系统。他引用 Physical Intelligence 联合创始人 Sergey Levine 在机器人领域的同类判断：把问题做大，让通用系统跨任务学习。

https://x.com/adityaag/status/2113666947766169868

Builder Zara Zhang

Zara Zhang is hosting a beginner-friendly event in San Mateo on November 7 where participants can build or rebuild a personal website alongside other makers. No coding or design experience is required.

Builder Zara Zhang 将于 11 月 7 日在 San Mateo 举办面向初学者的个人网站共创活动，参与者可以和其他创作者一起搭建或重做个人网站，无需编程或设计经验。

https://x.com/zarazhangrui/status/2113676956885807406

PODCASTS

No Priors: Democratizing Security for the Agentic Web with Cloudflare CTO Dane Knecht

The Takeaway: AI is compressing years of cybersecurity work into an 18-month emergency, but the same agents creating the threat can ultimately make software safer and rebuild the web's economic rails.

Cloudflare CTO Dane Knecht expects security to consume a majority of engineering attention during the near-term transition. AI has made sophisticated vulnerability discovery broadly accessible, turning incidents once seen annually into something closer to a weekly rhythm. His first-principles response is simple: reduce attack surface, simplify code, use safer runtimes, and build adversarial review into every stage of the agent development lifecycle. Cloudflare codifies its own engineering principles so agents can apply them during implementation and review, while humans retain judgment over architecture and complexity. As Knecht puts it, "The easiest thing you can do is to give less places for there to be vulnerabilities."

The larger shift is economic as well as technical. Humans and agents currently share a web designed for human browsing, forcing agents to waste tokens parsing pages while publishers absorb crawler costs. Knecht expects separate delivery paths for humans and machines, paired with discovery signals and micropayment rails. Cloudflare is already working on content monetization, agent wallets, stablecoin settlement, and a Signals project with OpenAI. His bet is that agents may finally make HTTP 402-style internet payments practical, allowing machines to pay creators at massive transaction volume instead of extracting content for free.

核心结论： AI 正把多年的网络安全工作压缩成一场持续约 18 个月的紧急转型，但制造威胁的同一类 agent，最终也可能让软件更安全，并重建 Web 的经济基础设施。

Cloudflare CTO Dane Knecht 预计，在近期转型阶段，安全工作将占据工程团队的大部分精力。AI 正把高级漏洞发现能力普及给更多人，使过去一年一遇的严重事件逐渐接近每周发生。他从第一性原理给出的应对方式很直接：缩小攻击面、简化代码、采用更安全的 runtime，并把对抗性审查嵌入 agent 开发生命周期的每个环节。Cloudflare 已将内部工程原则系统化，让 agent 在实现和 review 中持续执行，同时由人类把控架构与复杂度。正如 Knecht 所说：“最容易做到的事，就是减少可能存在漏洞的位置。”

更大的变化同时发生在经济层。人类与 agent 目前共用一套为人类浏览设计的 Web，agent 浪费 token 解析网页，内容方则承担爬取成本。Knecht 预计，未来将出现面向人类和机器的不同内容交付路径，并配套内容发现信号与微支付基础设施。Cloudflare 已在推进内容变现、agent wallet、stablecoin 结算，以及与 OpenAI 合作的 Signals 项目。他判断，agent 可能最终让基于 HTTP 402 的互联网支付真正落地，使机器能够以超高频交易向内容创作者付费，而不是免费抽取内容。

https://www.youtube.com/@NoPriorsPodcast

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