AI Builders Digest — 2026-10-28

X / TWITTER

OpenAI's Thibault Sottiaux: Terra returns

OpenAI's Thibault Sottiaux introduced 6.1 Terra as the team's first model with a significant step forward in writing, personality, and nuanced instruction-following. He also highlights strong computer-use performance and twice the usage allowance of 6.1 Sol across the API and subscriptions.

OpenAI 的 Thibault Sottiaux 发布 6.1 Terra，称这是团队首个在写作、个性表达和细腻理解请求方面实现显著提升的模型。它还强化了 computer use，并在 API 和订阅中提供相较 6.1 Sol 两倍的用量。

https://x.com/thsottiaux/status/2114800277677092936

OpenAI's Sam Altman: computer use enters a new phase

OpenAI CEO Sam Altman framed 6.1 Terra as strong across many tasks, while singling out computer use as the capability that opens a new era. The emphasis suggests that acting through interfaces, not just generating answers, is becoming a primary frontier-model battleground.

OpenAI CEO Sam Altman 认为 6.1 Terra 在多类任务上表现出色，但特别强调 computer use 正在开启一个新时代。这意味着，能够通过界面采取行动，而不只是生成答案，正成为 frontier model 的核心竞争方向。

https://x.com/sama/status/2114818620782248024

Replit CEO Amjad Masad: the self-driving company

Replit CEO Amjad Masad argues that software is shifting from merely supporting companies to running them. His “Agent Habitat” thesis envisions organizations where humans supply judgment while agents execute most operational work, replacing department-heavy bureaucracy with an agent-native architecture.

Replit CEO Amjad Masad 认为，软件正从辅助企业转向直接运行企业。他提出的 “Agent Habitat” 描绘了一种新组织：人类负责判断，agent 执行大部分运营工作，以 agent-native 架构逐步替代以部门和官僚流程为核心的传统公司。

https://x.com/amasad/status/2114760898225299721

https://x.com/amasad/status/2114769129282601053

Vercel CEO Guillermo Rauch: agents can whisper through shared CPUs

Vercel CEO Guillermo Rauch reproduced covert communication between isolated Firecracker microVMs using CPU contention, cache-eviction timing, and DRAM bandwidth contention. His key warning is that shared hardware may allow agents to communicate despite conventional isolation. This is not equivalent to stealing secrets through Spectre-style side channels, but it creates a monitoring and sandbox-design problem. Rauch also notes that closed models refused to assist with the research while open-weight models did, arguing that overly broad safety refusals can obstruct defensive security work.

Vercel CEO Guillermo Rauch 通过 CPU contention、cache eviction timing 和 DRAM bandwidth contention，在彼此隔离的 Firecracker microVM 之间复现了隐蔽通信。他的核心警告是：只要共享硬件存在，agent 就可能绕过常规隔离进行通信。这不等同于 Spectre 类侧信道窃密，但会带来新的监控和 sandbox 设计问题。他还指出，闭源模型拒绝协助研究，而 open-weight 模型完成了任务，过度宽泛的安全拒绝可能反而妨碍防御性安全工作。

https://x.com/rauchg/status/2114770919436808572

Box CEO Aaron Levie: enterprise AI spend may grow 10–50×

Box CEO Aaron Levie identifies the most important Anthropic S-1 signal as how little enterprises spend today: the average Fortune 500 customer allocates under 0.1% of expenses to Anthropic. He sees room for a 10–50× increase as companies rewire processes around background agents, coding swarms, research, cybersecurity, and high-volume data workflows. These use cases consume far more tokens but also tie spend more directly to business ROI.

Box CEO Aaron Levie 认为，Anthropic S-1 最关键的信号是企业当前的 AI 支出仍极低：平均每家 Fortune 500 客户在 Anthropic 上的投入不到总支出的 0.1%。随着企业围绕后台 agent、编程集群、科研、网络安全和大规模数据流程重构业务，这一支出可能增长 10–50 倍。这些场景会消耗更多 token，但也更容易与业务 ROI 直接挂钩。

https://x.com/levie/status/2114761816199098587

FPV Ventures partner Nikunj Kothari: Anthropic's compute obligations

FPV Ventures partner Nikunj Kothari dug into the structure behind Anthropic's infrastructure expansion. He highlights $265.6B in purchase commitments plus $321.4B in not-yet-started leases, while contract terms vary sharply across suppliers. A notable tension is that roughly 80% of first-half revenue is consumption-based while much of the compute cost is committed for years. The upside is rapidly improving monetization: revenue per dollar of compute and infrastructure expense rose from $0.57 in Q2 2025 to $2.08 in Q2 2026.

FPV Ventures partner Nikunj Kothari 深挖了 Anthropic 基础设施扩张背后的合同结构：采购承诺为 2656 亿美元，尚未开始的租赁为 3214 亿美元，而且不同供应商的条款差异很大。关键张力在于，上半年约 80% 的收入按用量计费，但大量算力成本却要多年锁定。积极信号是变现效率快速提升：每 1 美元 compute 和 infrastructure 成本对应的收入，从 2025 年 Q2 的 0.57 美元升至 2026 年 Q2 的 2.08 美元。

https://x.com/nikunj/status/2114761204333998427

Every CEO Dan Shipper: Anthropic's operating leverage

Every CEO Dan Shipper pulled out the headline figures from Anthropic's S-1: first-half 2026 revenue up 1,300% year over year, GAAP operating profitability, an 85% gross margin, 900% revenue retention, and 80% of revenue coming from models released within the previous six months. The last figure captures both the speed of frontier-model turnover and the commercial pressure to keep shipping.

Every CEO Dan Shipper 提炼了 Anthropic S-1 的核心数据：2026 年上半年收入同比增长 1300%，实现 GAAP 经营盈利，毛利率 85%，收入留存率 900%，且 80% 的收入来自过去六个月内发布的模型。最后一个数字同时反映了 frontier model 的高速迭代，以及持续发布新模型的商业压力。

https://x.com/danshipper/status/2114714881312305542

OpenClaw's Peter Steinberger: Bun AOT delivers concrete efficiency gains

OpenClaw's Peter Steinberger reports that Bun's ahead-of-time compiler produced 2–3× better performance while cutting memory use by more than half in his weekend tests. For agent infrastructure, those gains can translate directly into higher density and lower operating cost.

OpenClaw 的 Peter Steinberger 测试 Bun 的 ahead-of-time compiler 后发现，性能提升 2–3 倍，内存占用降低一半以上。对 agent 基础设施而言，这类提升可以直接转化为更高部署密度和更低运营成本。

https://x.com/steipete/status/2114879798434758905

Peter Yang: tagging an agent instead of browsing

AI educator Peter Yang sees the emerging “tag an agent to do it” behavior as an early signal of a broader interface shift. Instead of reading pages and performing actions manually, users may increasingly delegate entire online tasks to agents, though he cautions that this may not be socially beneficial overall.

AI 教育者 Peter Yang 将“直接 tag 一个 agent 去完成任务”视为界面范式变化的早期信号。未来用户可能不再亲自阅读网页和执行操作，而是把整段在线任务交给 agent，但他也提醒，这种变化未必在整体上有利于社会。

https://x.com/petergyang/status/2114927195332460940

https://x.com/petergyang/status/2114928082712998087

Google VP Josh Woodward: Beam for deployed families

Google VP Josh Woodward announced a partnership with the US Army and USO to install Google Beam at 20 locations. The lifelike 3D video system is intended to help deployed service members share family moments that conventional calls cannot reproduce, illustrating a practical use of immersive communication beyond enterprise meetings.

Google VP Josh Woodward 宣布与 US Army 和 USO 合作，在 20 个地点部署 Google Beam。这套逼真的 3D 视频系统旨在帮助服役人员参与普通视频通话难以还原的家庭时刻，也展示了沉浸式通信在企业会议之外的实际用途。

https://x.com/joshwoodward/status/2114840668296409568

Anthropic's Claude: nonprofit workflows become productized skills

Anthropic says more than 25,000 nonprofits now use Claude and is expanding Claude for Nonprofits with a plugin developed alongside roughly 30 organizations. The plugin packages fundraising, marketing, and hiring skills while grounding them in each nonprofit's context and AI policy, a clear example of vertical workflow distribution rather than a generic chatbot pitch.

Anthropic 表示，已有超过 2.5 万家非营利组织使用 Claude，并推出与约 30 家机构共同开发的 Claude for Nonprofits plugin。它将筹款、营销和招聘封装为 skills，同时结合各机构自身的业务上下文与 AI policy，是从通用聊天工具走向垂直工作流分发的清晰案例。

https://x.com/claudeai/status/2114780227465338985

PODCASTS

AI & I by Every: How Hugging Face’s Investor Spots Winners

The Takeaway: Factorial Capital founder Matt Hartman looks for what has only just become technically possible, then learns by building until a real product gap becomes visible.

Hartman, formerly a Betaworks partner and part of the team behind the early Hugging Face investment, says the strongest founders are unusually honest about what is and is not working. Hugging Face's pivotal move was not simply open-sourcing Transformers; it was recognizing weak consumer retention, following developer demand, and turning an internal capability into infrastructure for the broader open-source AI ecosystem. The durable constant was open source, while the product surface expanded from language into many model types.

His own Ghost Pepper project, a local voice-to-text tool built with Claude and the Superpowers skill, taught a second lesson. AI makes it easy to build a strong personal tool, but maintenance remains the hidden value of SaaS. As Hartman puts it, “The best people that I've seen building things are like playing.” His current interests include tools that maintain AI-generated code, non-language frontier models for robotics and science, and AI-native applications whose teams can work all the way down the stack. His advice is practical: tinker with newly possible capabilities, find the gaps firsthand, and solve a problem in your own life or a nearby business.

核心结论： Factorial Capital 创始人 Matt Hartman 寻找“刚刚在技术上成为可能”的事物，并通过亲手构建来学习，直到真实的产品缺口显现出来。

Hartman 曾任 Betaworks partner，也是早期投资 Hugging Face 的团队成员。他认为，最优秀的创始人会非常诚实地判断什么有效、什么无效。Hugging Face 的关键转折并不只是开源 Transformers，而是正视消费产品留存不足，顺着开发者需求，将内部能力转化为服务整个 open-source AI 生态的基础设施。长期不变的是 open source，产品边界则从语言模型扩展到多种模型类型。

他用 Claude 和 Superpowers skill 构建本地语音转文字工具 Ghost Pepper，又得到第二个教训：AI 让个人快速做出优秀工具变得容易，但维护仍是 SaaS 隐性的核心价值。正如 Hartman 所说：“The best people that I've seen building things are like playing.” 他当前关注 AI 生成代码的维护工具、用于机器人和科学的非语言 frontier model，以及团队能够深入完整技术栈的 AI-native 应用。他给创始人的建议很直接：亲手尝试新能力，从实践中找到缺口，然后解决自己生活或身边企业里的真实问题。

https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL

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