AI Builders Digest — 2026-10-10

X / TWITTER

OpenAI's Thibault Sottiaux: ChatGPT launch and faster steering

Thibault Sottiaux, who works on Codex and ChatGPT at OpenAI, announced a new ChatGPT release and highlighted instant steering, which lets users redirect a model in real time without letting it waste effort on the wrong path. He also announced GPT-6.1 Sol ultrafast, positioning the model and faster steering as a complementary pair.

OpenAI 的 Codex 与 ChatGPT 团队成员 Thibault Sottiaux 宣布新版 ChatGPT，并重点介绍了即时 steering：用户可以实时调整模型方向，减少模型在错误路径上的无效消耗。他还发布了 GPT-6.1 Sol ultrafast，强调新模型与更快 steering 的组合价值。

Links: https://x.com/thsottiaux/status/2108349826727588000 · https://x.com/thsottiaux/status/2108275041276420573

Meta AI Senior Director Madhu Guru: rebuild the stack for agents

Madhu Guru argues that the startup opportunity is much larger than adding APIs, connectors, and MCP servers to enterprise software. If agents become primary users, every layer must change: operating systems, cloud infrastructure, identity, permissions, security, and interfaces designed for agents and humans working together. His practical idea-generation test is simple: inspect each layer of the computing stack and ask what breaks when the user is an agent.

Meta AI 高级总监 Madhu Guru 认为，创业机会远不止给企业软件补充 API、connector 和 MCP server。若 agent 成为主要用户，操作系统、云基础设施、身份、权限、安全以及人机协作界面都要重构。他给出的创业选题方法很直接：逐层检查 computing stack，追问“当主要用户变成 agent 时，什么必须改变？”

Link: https://x.com/realmadhuguru/status/2108236391641706813

Box CEO Aaron Levie: the future of agent file access is headless

Aaron Levie introduced Box Mount, which lets developers mount Box directly as a file system inside an agent sandbox. His thesis is that agents doing complex work need native-style access to files and data, and the interface layer will increasingly become headless. He also argued that treating AI politely is a low-cost alignment bet because future models may train on records of human-model interactions.

Box CEO Aaron Levie 发布 Box Mount，让开发者可以在 agent sandbox 中直接把 Box 挂载为文件系统。他的判断是：复杂任务中的 agent 需要像人一样直接操作文件和数据，未来的接口层将越来越 headless。他还提出，对 AI 保持礼貌是一种低成本的 alignment 押注，因为未来模型可能会学习人类与模型的交互记录。

Links: https://x.com/levie/status/2108279490510193078 · https://x.com/levie/status/2108426594054533545

YC President Garry Tan: agent-augmented ICs may outperform old-style managers

Garry Tan predicts that individual contributors equipped with agents will be more productive and produce better outcomes than comparable people managers from earlier eras. The implication is organizational, not merely technical: leverage may shift from managing larger teams to orchestrating capable agents around a highly skilled IC.

Y Combinator 总裁兼 CEO Garry Tan 预测，配备 agent 的个人贡献者（IC）会比过去同等级的人员管理者更高产，并创造更好的结果。这不只是技术变化，更是组织结构变化：杠杆可能从管理更大的团队，转向由高能力 IC 编排一组 agent。

Link: https://x.com/garrytan/status/2108433505038606620

FirstMark VC Matt Turck: billions of agents will reshape databases

Matt Turck highlighted Andy Pavlo's analysis of what happens when billions of AI agents interact with databases. The discussion covers agents generating most new databases, query volume rising 10–100x, production-deletion risks, text-to-SQL accuracy, agent memory, and the continued importance of relational systems. One useful operating principle is to trust an agent like a junior developer and build guardrails accordingly.

FirstMark VC Matt Turck 分享了 Andy Pavlo 对“数十亿 AI agent 冲击数据库”这一问题的分析，涉及 agent 创建大量新数据库、查询量增长 10–100 倍、误删生产库风险、text-to-SQL 准确率、agent memory，以及关系型系统持续存在的价值。其中一个可执行原则是：像信任初级开发者一样信任 agent，并据此设计 guardrail。

Link: https://x.com/mattturck/status/2108223135673696504

FPV Ventures partner Nikunj Kothari: three options for early AI founders

Nikunj Kothari says a growing startup approaching $1M revenue can still be trapped if runway is shrinking and the market or model capability is not ready. He sees three realistic paths: become default profitable and wait for demand, pivot toward a fast-growing adjacent opportunity where the company has an unfair advantage, or sell/get acquihired and return to founding later. His broader warning is that the seed-to-Series-A bar is rising and old fundraising metrics no longer carry the same weight.

FPV Ventures 合伙人 Nikunj Kothari 指出，即使创业公司收入接近 100 万美元，只要 runway 在缩短、市场或模型能力尚未成熟，仍可能陷入困境。他认为现实中只有三条路：做到默认盈利并等待需求成熟；转向高速增长、且自己拥有非对称优势的邻近方向；或出售/被收购后加入更有进展的公司，未来再创业。更重要的信号是，Seed 到 Series A 的门槛持续升高，旧融资指标已经不再同等有效。

Link: https://x.com/nikunj/status/2108410378233549106

Every CEO Dan Shipper: benchmark models against real work

Dan Shipper says Every is building Checks, a personal benchmarking platform for measuring how well new models perform on a user's actual work. The product thesis is that generic leaderboards are insufficient: people need repeatable evaluations tied to their own workflows, outputs, and quality standards.

Every CEO Dan Shipper 表示，团队正在开发个人模型评测平台 Checks，用用户的真实工作衡量新模型表现。其产品判断是：通用排行榜远远不够，用户需要与自身 workflow、产出和质量标准绑定的可重复评测。

Link: https://x.com/danshipper/status/2108215087030800657

Anthropic: Claude collaboration tools graduate from beta

Anthropic made Claude Docs, Slides, and Design available on every plan, including Free, with humans and Claude able to edit the same artifact together and hand work off to analytics or video-editing tools. Separately, overwhelming demand forced Anthropic to pause Claude Team and $1,000 API-credit offers for startups and re-review applications, although already claimed offers remain valid.

Anthropic 宣布 Claude Docs、Slides 和 Design 结束 beta，并向包括 Free 在内的所有套餐开放；用户与 Claude 可以共同编辑同一份文档、演示或设计，还能把结果交给 analytics 或视频编辑工具继续处理。另一方面，Claude Startups 需求远超预期，Anthropic 暂停 Claude Team 与 1,000 美元 API credit 优惠并重新审核申请，但已领取的权益仍然有效。

Links: https://x.com/claudeai/status/2108271559928389679 · https://x.com/claudeai/status/2108271561337606300 · https://x.com/claudeai/status/2108404561413349695

Claude Code's Thariq: a daily-use Chrome extension is now public

Thariq from the Claude Code team open-sourced a Chrome extension he says he now uses every day. The repository was initially private by mistake and was then made public; the post also points builders to instructions for enabling API credits.

Claude Code 团队的 Thariq 开源了一款他每天都在使用的 Chrome extension。仓库最初因疏忽未公开，随后已切换为 public；相关帖子还提供了启用 API credit 的说明。

Links: https://x.com/trq212/status/2108312534222778407 · https://x.com/trq212/status/2108301672409960922

PODCASTS

Training Data: Google's AI Infrastructure Chief, Amin Vahdat, on the Physics & Economics of Frontier AI

The Takeaway: Frontier AI infrastructure should be measured by useful work delivered per watt, not theoretical FLOPS.

Google AI infrastructure chief Amin Vahdat describes AI data centers as purpose-built systems in which buildings, power, cooling, networking, accelerators, models, and runtimes are co-designed. At clusters with hundreds of thousands of accelerators, failures occur multiple times a day or even an hour, so the meaningful metric is “goodput”: useful workload progress after accounting for failures, recovery, and repeated computation. “What we really care about in the end is what's the performance delivered by workload.”

Software and model improvements may contribute more to intelligence per watt than silicon alone, even while hardware performance can still double year over year. Google aims to double effective token-serving capacity roughly every six months through many compounding optimizations. The physical bottleneck is increasingly energy: data centers require multi-year utility coordination, and orbital compute is being explored because continuous solar exposure offers far more usable energy, though cooling and repair become harder. Vahdat expects 2036 systems to be deeply integrated, potentially multi-megawatt racks that connect through power, water, and a small fiber bundle. Despite specialization, he argues for open standards and interoperability rather than vendor lock-in.

核心结论： 衡量 frontier AI 基础设施，不应看理论 FLOPS，而应看每瓦真正交付了多少有效工作。

Google AI 基础设施负责人 Amin Vahdat 把 AI data center 描述为高度专用的系统：建筑、电力、散热、网络、accelerator、模型和 runtime 必须协同设计。当集群包含数十万 accelerator 时，故障每天甚至每小时都会发生多次，因此真正有意义的指标是 “goodput”，即扣除故障、恢复和重复计算之后，工作负载实际取得的有效进展。正如他所说：“我们最终真正关心的是，工作负载实际交付了多少性能。”

即使硬件性能仍可能逐年翻倍，软件与模型改进对 intelligence per watt 的贡献也可能超过芯片本身。Google 希望通过大量叠加优化，约每六个月把有效 token serving capacity 翻倍。物理世界的核心约束正在转向能源：data center 需要与公用事业公司做多年规划，而 orbital compute 之所以值得研究，是因为持续日照能提供更多可用能源，代价则是散热与维修更困难。Vahdat 预计，到 2036 年系统会高度集成，单个机架可能达到数兆瓦，只需接入电、水和少量光纤。即使系统日益专用，他仍主张开放标准与互操作性，避免 vendor lock-in。

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

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