AI Builders Digest — 2026-09-15

X / TWITTER

Boris Cherny, Claude Code at Anthropic

Boris Cherny highlighted Fable's use of Claude to solve the Cyphral Distich, a 370-year-old cipher. It is a striking example of an LLM being used not merely to retrieve known facts, but to support original historical cryptanalysis.

Boris Cherny 分享了 Fable 使用 Claude 破解拥有 370 年历史的 Cyphral Distich 密码。这是一个值得关注的案例：LLM 不只是检索已知事实，而是在辅助完成原创性的历史密码分析。

https://x.com/bcherny/status/2099322487603634395

Peter Yang, AI educator and interviewer

Peter Yang cited Brex CEO Pedro Franceschi's estimate that 84% of the world has never used AI, about 0.3% pays for an AI chatbot, and only 0.04% uses agents effectively. Even if the February 2026 figures have improved, his point is that practical agent adoption remains at the very beginning.

AI 教育者 Peter Yang 引用 Brex CEO Pedro Franceschi 的估算：全球 84% 的人从未使用过 AI，约 0.3% 的人付费使用 AI chatbot，真正有效使用 agent 的人只有 0.04%。即使这组 2026 年 2 月的数据已有增长，核心判断仍是：agent 的实际普及才刚刚开始。

https://x.com/petergyang/status/2099200231820963891

Replit CEO Amjad Masad

Amjad Masad said AI coding had priced too many users out, but that it is now free to build again. The substantive signal is renewed competition around zero-cost entry points for AI-assisted software creation.

Replit CEO Amjad Masad 表示，AI coding 一度让太多用户因价格被挡在门外，但现在又可以免费构建软件了。真正的行业信号是：AI 辅助开发正在重新围绕零成本入口展开竞争。

https://x.com/amasad/status/2099197117013340450

Box CEO Aaron Levie

Aaron Levie argued that “pacing” AI development should not mean an arbitrary slowdown or regulation designed to hobble competitors. Because AI will underpin trading, medical devices, biotech, defense, and government systems, he sees safety and alignment as necessary engineering requirements, with the hard problem being how to achieve them without suppressing innovation or competition.

Box CEO Aaron Levie 认为，AI development 的“pacing”不应等同于任意减速，也不应成为限制竞争者的监管工具。由于 AI 将成为交易、医疗设备、生物科技、国防和政府系统的基础设施，安全与 alignment 是必要的工程要求；真正困难的是如何在不压制创新和竞争的前提下实现它们。

https://x.com/levie/status/2099167992835924301

Builder Zara Zhang

Zara Zhang identified a concrete agent UX failure in Astra: after accepting a user's correction, the model says what it should do but does not execute the corrected action. Her complaint captures the gap between conversational agreement and reliable agency.

Builder Zara Zhang 指出了 Astra 的一个具体 agent UX 问题：模型接受用户纠正后，只复述“应该怎么做”，却没有真正执行修正后的动作。这个问题揭示了“语言上同意”与“可靠行动能力”之间的差距。

https://x.com/zarazhangrui/status/2099348631291883945

FPV Ventures partner Nikunj Kothari

Nikunj Kothari warned startup candidates that a higher entry valuation and higher total compensation can be bugs rather than benefits. Candidates should independently assess market size, traction, positioning, plausible exit value, 409A pricing, and taxes, especially when a company trades at extreme multiples such as 100x ARR.

FPV Ventures partner Nikunj Kothari 提醒求职者，更高的入职估值和总薪酬可能不是优势，反而是风险。候选人应独立评估市场、traction、竞争位置、合理退出估值、409A 定价和税务影响，尤其要警惕 100 倍 ARR 这类极端估值。

https://x.com/nikunj/status/2099198567923765357

Peter Steinberger, OpenClaw and OpenAI builder

Peter Steinberger said an upcoming release or dev channel will make worktrees roughly 80% faster by using folder cloning on APFS, Btrfs, XFS, and ReFS, while also reducing disk usage. This is a practical infrastructure improvement for coding agents that create isolated worktrees frequently.

OpenClaw 与 OpenAI builder Peter Steinberger 表示，下一版本或 dev channel 将利用 APFS、Btrfs、XFS 和 ReFS 的目录克隆能力，让 worktree 创建速度提升约 80%，同时减少磁盘占用。这是对频繁创建隔离 worktree 的 coding agent 很实用的基础设施优化。

https://x.com/steipete/status/2099197266636783989

OpenAI CEO Sam Altman

Sam Altman framed two major AI risks: humanity losing control to AI, and excessive power concentrating in one person, company, lab, or country. He supports consistent federal frontier-safety requirements, independent auditing, pre-training safety cases for major reinforcement-learning runs, and shared standards for misalignment and monitoring. His definition of “pacing” is slower progress where safety demands it, not stopping progress.

OpenAI CEO Sam Altman 将 AI 的主要风险归纳为两类：人类失去对 AI 的控制，以及权力过度集中于某个人、公司、实验室或国家。他支持统一的联邦 frontier safety 要求、独立审计、在重大 reinforcement learning 训练前建立 safety case，以及形成针对 misalignment 和 monitoring 的行业共同标准。他所说的“pacing”是在安全需要时适度放慢，而不是停止进步。

https://x.com/sama/status/2099352016988614852

https://x.com/sama/status/2099348812305473766

PODCASTS

No Priors: Redefining Chip Architecture with Arm CEO Rene Haas

The Takeaway: AI demand is turning Arm from an IP supplier into a broader systems company, while making CPUs, supply-chain access, and power efficiency more strategically important, not less.

Arm CEO Rene Haas says chip design still takes 24 to 36 months, with verification, validation, debugging, and documentation consuming more time than the architecture itself. AI already helps heavily in those stages, and 80% to 90% of Arm engineers use it daily. “The genie's out of the bottle and there's no stopping that.” He believes simpler designs could eventually move from an idea to a fabrication-ready GDSII file, although frontier physical design remains constrained by proprietary training data.

Arm's move from component IP to compute subsystems and now physical CPUs reflects customers' need for faster time-to-market. Haas expects wafer, memory, packaging, labor, and data-center construction constraints to persist for three to five years. He also argues that accelerators do not replace CPUs: GPUs may produce tokens, but CPUs still orchestrate where those tokens go. Robotics is another major opportunity because sensing, control, and many current robot “brains” already rely on Arm architectures.

核心结论： AI 需求正在推动 Arm 从 IP 供应商走向更广泛的系统公司，同时让 CPU、供应链获取能力和能效变得更加重要，而不是被弱化。

Arm CEO Rene Haas 表示，芯片设计周期仍需 24 至 36 个月，其中 verification、validation、debug 和文档工作比架构设计本身更耗时。AI 已经深度参与这些环节，Arm 内部 80% 至 90% 的工程师每天都在使用 AI。“精灵已经从瓶子里出来了，没有办法再阻止它。”他判断，较简单的芯片未来可能从概念直接生成可送交晶圆厂的 GDSII 文件，但先进 physical design 仍受制于专有训练数据不足。

Arm 从组件 IP 扩展到 compute subsystem，再进入实体 CPU，根本原因是客户需要更快的上市速度。Haas 预计晶圆、内存、先进封装、劳动力和数据中心建设的约束还会持续三到五年。他也强调 accelerator 不会取代 CPU：GPU 可以生成 token，但 token 如何被调度和送达仍需要 CPU。Robotics 也是 Arm 的重要机会，因为传感、控制以及许多现有机器人的“大脑”已经运行在 Arm 架构上。

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

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