AI Builders Digest — 2026-10-24

2026-10-24

AI Builders Digest — 2026-10-24

X / TWITTER

Swyx (AI Engineer / Latent Space)

Swyx says accurate time estimation is becoming a core agent capability: Opus 5.5 is unusually well calibrated about completing in hours what humans estimate in days. His operational advice is equally direct: when AI uncovers bugs, do not blame the tool; focus the team on clearing the defects quickly, while letting agents estimate and repair their own pipelines during execution.

Swyx 认为，准确估时正成为 agent 的核心能力：Opus 5.5 能较好地区分「人类需要数天」与「并行 agent 只需数小时」的任务。他的工程建议也很直接：AI 找出 bug 后，不要怪工具，团队应立即集中精力清障，并让 agent 在执行过程中持续估算和修复自己的 pipeline。

Boris Cherny (Claude at Anthropic)

Boris Cherny announced that Anthropic is donating Agent Skills, which began as a Claude Code experiment and has since gained industry adoption, to the Agentic AI Foundation alongside MCP. The move positions Skills as shared agent infrastructure rather than a Claude-only feature.

Boris Cherny 宣布，Anthropic 将 Agent Skills 捐赠给 Agentic AI Foundation，与 MCP 一同进入开放治理。Agent Skills 最初是 Claude Code 的实验，如今已被行业广泛采用，这次转移意味着它正在从 Claude 专属功能变为 agent 生态的公共基础设施。

https://x.com/bcherny/status/2113333859907747871

Thibault Sottiaux (Codex & ChatGPT at OpenAI)

Thibault Sottiaux reported that the Decisions API reached GA, OpenResponses joined the Linux Foundation, Codex received performance improvements, and Uber is going all-in on Codex for work. He also teased another small release scheduled before 10 a.m. Pacific the following day.

OpenAI Codex 与 ChatGPT 团队的 Thibault Sottiaux 表示，Decisions API 已 GA，OpenResponses 加入 Linux Foundation，Codex 获得性能改进，Uber 也在全面采用 Codex 工作流。他还预告次日太平洋时间上午 10 点前会有一次小型发布。

Peter Yang (AI educator and product builder)

Peter Yang highlighted the gap between frontier technology and real operations: while AI breakthroughs accelerate, cancer-care workflows still depend on faxed documents and hand-delivered CD-ROM imaging. He also published a six-step account-security workflow with a prompt that uses AI to guide users and generate a personalized security report.

AI 教育者与产品创作者 Peter Yang 指出了前沿技术与真实运营之间的断层：AI 突飞猛进，但癌症治疗流程仍依赖传真文件和人工递送影像 CD-ROM。他还发布了一套六步账户安全流程，并提供 prompt，让 AI 引导用户完成检查并生成个性化安全报告。

Google Labs

Google Labs opened Dreambeans to U.S. users aged 18 and over. The experiment avoids streaks and doomscrolling, instead using personalized prompts to encourage people to leave their phones and explore the physical world, with more regions planned.

Google Labs 面向美国 18 岁以上用户开放 Dreambeans。这个实验刻意避开连续签到和 doomscrolling，改用个性化提示鼓励用户放下手机、探索现实世界，后续将扩展到更多地区。

https://x.com/GoogleLabs/status/2113340712523841696

Amjad Masad (Replit CEO)

Replit CEO Amjad Masad shared how Replit is protecting customer applications from the recent wave of vulnerabilities, signaling that security hardening is becoming a product-level responsibility for AI app platforms rather than an optional developer task.

Replit CEO Amjad Masad 分享了 Replit 如何保护用户应用免受近期一系列漏洞影响。这释放出一个信号：对于 AI 应用平台，安全加固正在从开发者的可选任务变成平台自身必须承担的产品责任。

https://x.com/amasad/status/2113474587166171560

Guillermo Rauch (Vercel CEO)

Vercel CEO Guillermo Rauch argues that static credentials should be replaced by short-lived, self-destructing keys built around OIDC, much as the industry standardized TLS. He also highlighted Generative UI connected through AI Gateway, where adding a tool through AI SDK can expose capabilities such as search, weather, sports, stocks, and elections with minimal integration work.

Vercel CEO Guillermo Rauch 主张用基于 OIDC 的短期、自毁式密钥取代静态凭证，就像行业当年共同普及 TLS 一样。他还展示了通过 AI Gateway 连接的 Generative UI：开发者只需在 AI SDK 中增加极少代码，即可接入搜索、天气、体育、股票和选举等工具能力。

Alex Albert (Anthropic Research)

Anthropic researcher Alex Albert sees AI-assisted local businesses as a major product frontier. He suggests tools such as Claude Duet could eventually matter more to small operators than Square or Google Maps by helping run day-to-day workflows, not merely providing another software interface.

Anthropic 研究员 Alex Albert 认为，AI 辅助本地小企业会成为重要的产品前沿。他判断，Claude Duet 一类工具未来对小企业的影响可能超过 Square 或 Google Maps，因为它们不是再增加一个软件界面，而是直接参与日常经营流程。

Aaron Levie (Box CEO)

Box CEO Aaron Levie says security through obscurity is finished because autonomous systems can probe every crevice and attack permutation. Legacy enterprise infrastructure, access controls, and data pipelines will require years of modernization; defensive agents will help accelerate the work and simulate risk, but the scale of remediation means enterprise AI diffusion may take longer than optimists expect.

Box CEO Aaron Levie 认为，「通过隐藏实现安全」已经失效，因为自主系统可以遍历每个角落和攻击组合。传统企业基础设施、访问控制和数据 pipeline 需要多年现代化改造；防御型 agent 会加快修复并模拟风险，但治理规模也意味着企业 AI 的全面扩散可能比乐观者预期更慢。

https://x.com/levie/status/2113499900399034689

Garry Tan (Y Combinator President & CEO)

Y Combinator President and CEO Garry Tan endorsed Aside Browser as a practical way to make agents better at using the web. He also pointed to AI's deeper impact on established SaaS businesses, emphasizing operational transformation rather than surface-level feature additions.

Y Combinator 总裁兼 CEO Garry Tan 推荐 Aside Browser，认为它能切实提升 agent 使用 Web 的能力。他同时强调，AI 正在深入改造成熟 SaaS 业务，真正的变化是运营模式转型，而不只是增加表层 AI 功能。

Matt Turck (FirstMark VC and MAD Podcast host)

FirstMark investor Matt Turck says AI market boundaries are collapsing: chipmakers compete with inference providers, inference platforms with neo-clouds and post-training firms, models with apps and chips, and apps with models, while NVIDIA sells across the stack. He also shared a detailed look inside Mistral's open-model comeback, including a 3,800-GPU training run, mixture-of-experts scaling, native multimodality, RL, and the prospect of 10-trillion-parameter systems.

FirstMark 投资人 Matt Turck 认为，AI 市场边界正在消失：芯片商与 inference 服务竞争，inference 平台与 neo-cloud 和 post-training 公司竞争，模型同时与应用和芯片竞争，应用也开始反向竞争模型，而 NVIDIA 几乎向所有层级供货。他还分享了 Mistral 开放模型回归的深度内容，涉及 3,800 张 GPU 训练、Mixture of Experts 扩展、原生多模态、RL，以及未来 10 万亿参数系统。

Nikunj Kothari (FPV Ventures partner)

FPV Ventures partner Nikunj Kothari says the seed-to-Series-A squeeze is at an all-time high. Roughly 10% of seed companies he meets are acquired within months, capturing the current paradox: it has never been easier to start dreaming, but scaling a venture has rarely been harder.

FPV Ventures 合伙人 Nikunj Kothari 表示，Seed 到 Series A 的挤压已达到历史高位。他接触的 Seed 公司中约有 10% 会在数月内被收购，反映出当前创业的悖论：启动梦想从未如此容易，把公司真正规模化却极其困难。

https://x.com/nikunj/status/2113405034826477910

Aditya Agarwal (SPC General Partner and Bevel Health co-founder)

SPC General Partner Aditya Agarwal says robotics is following the same interface shift as LLMs: from fine-tuning, to prompting, to simply describing the desired outcome. Physical Intelligence co-founder Sergey Levine's π0.7 can learn new skills through plain-language coaching, suggesting natural language may become the practical programming layer for robots.

SPC General Partner、Bevel Health 联合创始人 Aditya Agarwal 认为，机器人正在经历与 LLM 相同的交互转变：从 fine-tuning，到 prompting，再到直接描述想要的结果。Physical Intelligence 联合创始人 Sergey Levine 的 π0.7 已能通过自然语言指导学习新技能，这意味着自然语言可能成为机器人的实用编程层。

Claude (Anthropic)

Anthropic announced same-day access on the native Claude Platform to Opus 5.5 fast mode, code execution, the Skills API, and other betas. Anthropic-managed Claude Models are now generally available on Google Cloud with IAM, billing, and commitment retirement, while Claude apps will default to auto mode for Pro, Max, and Team users, using automated checks to block risky actions instead of requesting approval at every step.

Anthropic 宣布，Claude 原生平台现可在功能上线当天使用 Opus 5.5 fast mode、代码执行、Skills API 等 beta。由 Anthropic 托管的 Claude Models 也已在 Google Cloud GA，支持 IAM、计费与承诺额度消耗；同时，Pro、Max 和 Team 用户的 Claude 应用将默认启用 auto mode，用自动化检查拦截高风险操作，而不是每一步都请求批准。

PODCASTS

Training Data: David Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience

The Takeaway: David Silver believes the route to superintelligence is not a larger language model, but an autonomous “superlearner” that keeps discovering knowledge from experience without human breadcrumbs.

AlphaGo pioneer David Silver and scaled-RL researcher Lasse Espeholt founded Ineffable Intelligence around a deliberately contrarian bet: systems trained primarily on human language may be the wrong vehicle for open-ended intelligence. Silver left DeepMind because this path deserved a lab wholly focused on experience-based learning, rather than a corner of an organization optimized around LLM products.

Their target is a learner with both depth and breadth: it should begin knowing almost nothing, continue beyond human knowledge, and work across simulated, digital, and physical environments. “Everything we learned, we discovered for ourselves,” Silver says, arguing that AI has not yet reproduced humanity's cumulative process of autonomous discovery.

The practical program combines a large-scale reinforcement-learning stack with “crucibles,” rich simulated environments where agents can rediscover milestones from mathematics, science, language, and technology. Crucially, the team does not lay a trail of tightly specified rewards toward known answers. It gives agents simple objectives and expects useful, surprising behaviors to emerge. The benchmark is not whether the system imitates human artifacts, but whether it independently invents tools and concepts because they help it succeed.

核心结论： David Silver 认为，通往 superintelligence 的关键不是更大的语言模型，而是一个无需人类铺设线索、能够持续从经验中自主发现知识的「superlearner」。

AlphaGo 先驱 David Silver 与大规模 RL 研究者 Lasse Espeholt 创办 Ineffable Intelligence，押注一条刻意逆势的路线：主要依靠人类语言训练的系统，可能并不是开放式智能的正确载体。Silver 离开 DeepMind，是因为这条路线值得由一个完全专注于经验学习的实验室推进，而不是依附在以 LLM 产品为中心的组织角落。

他们的目标同时追求深度与广度：系统应从几乎一无所知开始，持续探索到超越人类知识，并能适应模拟、数字和物理环境。Silver 说：「我们学会的一切，都是自己发现的。」他认为，AI 尚未复现人类通过集体探索不断积累知识的过程。

实际研发由两部分组成：大规模 reinforcement learning 技术栈，以及被称为「crucibles」的丰富模拟环境，agent 可以在其中重新发现数学、科学、语言与技术史上的关键里程碑。关键区别是，团队不会用高度定制的 reward 为已知答案铺路，只给出简单目标，期待有用且令人意外的行为自行涌现。真正的衡量标准不是系统能否模仿人类成果，而是它是否会因为解决问题的需要而独立发明工具与概念。

https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8

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