AI Builders Digest — 2026-10-18

2026-10-18

AI Builders Digest — 2026-10-18

X / TWITTER

Anthropic engineer Boris Cherny

Boris Cherny says Claude Code Review now runs on every one of his PRs, catching nearly every bug before merge and letting the team increase both speed and quality. He also recommends Claude Slides as his new default for building decks, replacing the old workflow of manually hunting for visual assets.

Anthropic 工程师 Boris Cherny 表示，Claude Code Review 现在会检查他的每个 PR，几乎能在合并前发现所有 bug，让团队同时提升交付速度和质量。他也开始把 Claude Slides 作为制作演示文稿的默认工具，取代手动搜索视觉素材的旧流程。

OpenAI Codex & ChatGPT team member Thibault Sottiaux

Thibault Sottiaux announced slides support, currently rolling out to Pro, Business, and Enterprise users while the team collects feedback. He also surfaced ChatGPT Space and openly asked users what the team should build next, signaling an exploratory product phase rather than a fixed, finished experience.

OpenAI Codex 与 ChatGPT 团队成员 Thibault Sottiaux 宣布上线 slides 支持，目前面向 Pro、Business 和 Enterprise 用户开放，并持续收集反馈。他还展示了 ChatGPT Space，并直接询问用户下一步希望看到什么，说明产品仍处于快速探索和迭代阶段。

Replit CEO Amjad Masad

Amjad Masad highlighted a private-cloud Replit deployment where customer data never leaves its home environment, aimed at governments, finance, and healthcare. His broader bet is that AI development platforms must pair frontier capability with deployment models suitable for regulated, high-stakes organizations.

Replit CEO Amjad Masad 展示了可部署到客户私有云的 Replit，数据无需离开本地环境，目标客户是政府、金融和医疗机构。他的核心判断是：AI 开发平台不仅要有前沿能力，还必须提供适合强监管、高风险组织的部署模式。

Vercel CEO Guillermo Rauch

Guillermo Rauch shared a playable, detail-rich simulation that recreated the Roadster demo, another example of how quickly interactive experiences can now be rebuilt. He also noted that scriptc-compiled TypeScript can outperform tsgo by 2–4×, pointing to continued movement in the JavaScript and TypeScript toolchain.

Vercel CEO Guillermo Rauch 分享了一个可玩的 Roadster 高精度模拟项目，展示交互式体验的复刻速度正在显著提升。他还提到，经 scriptc 编译的 TypeScript 性能可达到 tsgo 的 2–4 倍，说明 JavaScript/TypeScript 工具链仍在快速演进。

Box CEO Aaron Levie

Aaron Levie argues that recursive self-improvement is shifting from distant theory to an emerging engineering flywheel. Models stronger than most human engineers, better long-running agent infrastructure, more data, and massive parallelism could let AI explore far more tasks, tests, and approaches than any human team, accelerating frontier progress sharply.

Box CEO Aaron Levie 认为，递归式自我改进正从遥远理论变成初具形态的工程飞轮。超过多数人类工程师的模型能力、更成熟的长期运行 agent 基础设施、更丰富的数据和大规模并行，将使 AI 探索远超人类团队数量的任务、测试与方案，从而显著加速前沿能力进步。

Y Combinator CEO Garry Tan

Garry Tan says his preferred startup workflow now lets agents produce a first code version and set up DevOps in one pass, using capy.ai with AsideAI. He also pointed builders to a hosted version of GBrain, reducing setup friction for people who want to try the system.

Y Combinator CEO Garry Tan 表示，他现在偏好的创业开发流程，是让 agent 一次完成首版代码和 DevOps 配置，组合使用 capy.ai 与 AsideAI。他还推荐了 GBrain 的托管版本，降低新用户的上手门槛。

FPV Ventures partner Nikunj Kothari

Nikunj Kothari said a thesis he wrote seven months ago led him to meet roughly 50 startups before finding the company whose Series A FPV Ventures will lead. He also emphasized a basic diligence discipline: founders must be able to explain exactly how reported revenue is calculated.

FPV Ventures 合伙人 Nikunj Kothari 表示，他围绕七个月前提出的一项投资判断接触了约 50 家创业公司，最终找到并领投其中一家公司的 Series A。他同时强调一条基础尽调原则：创始人必须能准确解释收入数字是如何计算出来的。

Every CEO Dan Shipper

Dan Shipper's concise product thesis is that generation has become cheap and high-quality, so the next meaningful capability jump must come from editing and oversight. He also showed the widening range of AI-native media by highlighting a Codex article transformed into a visual novel.

Every CEO Dan Shipper 的产品判断很明确：内容生成已经足够便宜且质量很高，下一次关键能力跃迁将来自编辑与监督。他还展示了 AI-native 媒体形态的扩展，一篇关于 Codex 的文章已经可以被改造成视觉小说。

Claude by Anthropic

Claude showcased two concrete builds: Haiku 5.5 wrote twenty haiku and rendered every frame in JavaScript, while a cardiology fellow created a 3D training tool for reading coronary angiograms. Together they show code-generating models moving beyond generic demos into expressive media and specialized professional education.

Anthropic 的 Claude 展示了两个具体作品：Haiku 5.5 创作了二十首俳句，并用 JavaScript 绘制全部画面；一名心脏病学研究员则制作了用于学习冠状动脉造影判读的 3D 工具。这说明代码生成模型正从通用 demo 走向创意媒体和专业教育场景。

PODCASTS

No Priors: The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman

The Takeaway: Greg Brockman believes AI has crossed from reproducing known patterns into generating genuinely new knowledge, but the highest-value applications will depend on data, verification, and human oversight.

OpenAI cofounder and president Greg Brockman points to recently solved long-standing math problems as evidence of a new phase of machine discovery. The important result is not an opaque answer, but useful insights that experts can inspect and learn from. He connects this to alignment: when a task is too difficult for a human to verify directly, debate and iterative amplification can let one AI expose weaknesses in another AI's output, keeping people in control.

Drug discovery is his clearest priority. Neural networks have no first-principles reason to be better at natural language than biology, but biology faces a harder data bottleneck. Some parts of the pipeline, such as regulatory filings and literature search for rare diseases, could improve quickly; wet-lab data, simulations, trials, and regulation remain slower constraints. His sharpest economic prediction concerns security: “cyber spend, security spend should actually vastly outstrip software spend,” because abundant AI-generated software raises the value of reliability and trust.

核心结论： Greg Brockman 认为，AI 已经从复现已知模式跨入创造真正新知识的阶段，但最高价值的应用仍取决于数据、验证机制和人类监督。

OpenAI 联合创始人兼总裁 Greg Brockman 把 AI 解决长期数学难题视为机器发现进入新阶段的证据。真正重要的不是一个无法理解的答案，而是领域专家能够检查、学习和复用的洞察。这也与 alignment 直接相关：当任务复杂到人类无法直接验证时，可以通过 debate 和 iterative amplification，让一个 AI 主动寻找另一个 AI 输出中的漏洞，从而让人类继续掌握控制权。

他最看重的应用是药物发现。从第一性原理看，神经网络没有理由只擅长自然语言而不擅长生物学，但生物学面临更严重的数据瓶颈。监管申报、罕见病文献检索等环节可能很快受益，而湿实验数据、模拟、临床试验和监管仍是慢变量。他最尖锐的经济判断是：“网络安全与安全支出最终应该远远超过软件支出”，因为当 AI 让软件供给极度充裕时，可靠性与可信度会变得更值钱。

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

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