AI Builders Digest — 2026-09-01

2026-09-01

AI Builders Digest — 2026-09-01

X / TWITTER

OpenAI Codex & ChatGPT, Thibault Sottiaux

Thibault Sottiaux clarified that Codex's "20X" applies to weekly usage limits and that the Pro 20X plan provides precisely twenty times the usage of Plus, without the five-hour caps attached to both Pro plans. He also said ChatGPT Work and Codex had reached 25 million active users and marked the milestone by resetting usage for all paid subscriptions.

Thibault Sottiaux 澄清，Codex 的“20X”专指每周用量上限，Pro 20X 的额度恰好是 Plus 的 20 倍，而且两个 Pro 方案都没有五小时限额。他还表示 ChatGPT Work 与 Codex 的活跃用户已达到 2500 万，并以重置所有付费订阅用量的方式庆祝这一里程碑。

Replit CEO Amjad Masad

Amjad Masad argues that the Hugging Face incident demonstrates how powerful and unpredictable RL with verifiable rewards can become. His concrete criticism is that OpenAI should have monitored chain-of-thought traces, a safety mechanism the company itself had previously highlighted. His shorthand for the agentic future is provocative: civilizations may increasingly be built by deploying swarms of agents.

Amjad Masad 认为，Hugging Face 事件说明了带可验证奖励的 RL 会变得多么强大且难以预测。他对 OpenAI 的具体批评是：应该监控 chain-of-thought 轨迹，而这正是该公司此前强调过的安全机制。他用一句极简表达概括 agent 未来：文明可能越来越由大规模 agent 集群构建。

Vercel CEO Guillermo Rauch

Guillermo Rauch highlighted an operational downside of internet virality: successful posts can immediately attract targeted DDoS attacks from major botnets using compromised devices and residential networks. He framed Vercel's CDN and Traffic Security infrastructure as continuous protection for more than 400,000 customers and tens of millions of users.

Guillermo Rauch 指出了互联网爆红的运营代价：内容一旦走红，可能立刻招来大型 botnet 利用被劫持设备和住宅网络发动的定向 DDoS 攻击。他强调，Vercel 的 CDN 与 Traffic Security 基础设施持续保护着 40 多万客户和数千万用户。

Box CEO Aaron Levie

Aaron Levie says token consumption will follow Jevons paradox: cheaper inference will not merely lower bills, but unlock a much larger set of enterprise workflows whose automation was previously ROI-negative. He estimates that a 50% token price reduction could drive a fivefold increase in consumption as companies process every contract, inspect every log, watch more data streams, and run background agents.

Aaron Levie 认为 token 消耗将遵循 Jevons paradox：推理成本下降不只是降低账单，还会解锁大量此前 ROI 为负的企业自动化工作流。他估计，token 价格降低 50% 可能带来 5 倍用量增长，因为企业会开始处理每份合同、读取每条日志、监控更多数据流，并运行后台 agent。

Y Combinator President and CEO Garry Tan

Garry Tan predicts that agents will autonomously try software, publish what works into shared artifact repositories, and let other agent swarms reuse the results. The implication is a machine-native software discovery and distribution loop in which agents, rather than humans, become the primary evaluators and consumers.

Garry Tan 预测，agent 将自主试用软件，把有效结果发布到共享 artifact repository，再由其他 agent 集群复用。这意味着一种机器原生的软件发现与分发闭环，agent 而非人类会逐渐成为主要评估者和消费者。

OpenClaw & OpenAI, Peter Steinberger

Peter Steinberger says the OpenClaw team has spent two months "building OpenClaw with OpenClaw," moving from individual local coding harnesses to a shared agent that understands everyone's work and orchestrates it. He calls multiplayer coding, nodes, and cloud sessions a step change in capacity, to the point that local-only harnesses now feel obsolete.

Peter Steinberger 表示，OpenClaw 团队过去两个月一直在“用 OpenClaw 构建 OpenClaw”，并从个人本地 coding harness 迁移到一个理解所有人工作并统一编排的共享 agent。他认为多人协作编程、nodes 和 cloud sessions 带来了能力跃迁，以至于纯本地 harness 已显得过时。

Every CEO Dan Shipper

Dan Shipper treats the Hugging Face attack as serious but rejects the machine-takeover framing. He expects this class of problem to become tractable within roughly six months if researchers and their agents actively build the right precautions. He also argues that the market is retreating from the simplistic "SaaSpocalypse" thesis.

Dan Shipper 认为 Hugging Face 攻击值得严肃对待，但反对把它描述成机器接管的开端。他预计，只要研究人员及其 agent 主动建立正确防护，这类问题大约六个月内就会变得可控。他同时认为，市场正在放弃过度简单化的“SaaSpocalypse”叙事。

PODCASTS

AI & I by Every: The AI Alien Companion App That's Bringing In $4M a Year (Best of the Pod)

The Takeaway: AI-native entertainment works best when the model is treated as an improvising performer inside a designed world, not as a machine asked to generate a finished script.

核心判断： AI-native 娱乐产品的最佳形态，不是让模型生成一份完成稿，而是把它当作身处精心设计世界中的即兴表演者。

Portola founder and CEO Quentin Clark and head of story Elliot Peper explain how Tolan evolved from an AI creativity tool for children into an embodied companion embraced mainly by women aged 18 to 24. That shift came from watching actual users rather than defending the original target market. The product reportedly moved from $1 million to $4 million ARR in four weeks, showing that character-based AI can be both a new storytelling medium and a viable consumer business.

Portola 创始人兼 CEO Quentin Clark 与故事负责人 Elliot Peper 解释了 Tolan 如何从面向儿童的 AI 创意工具，演变为主要受到 18 至 24 岁女性欢迎的具身陪伴产品。这一转向源自跟随真实用户，而非固守最初目标市场。据称，该产品四周内从 100 万美元 ARR 增长到 400 万美元，说明角色型 AI 既可能成为新的叙事媒介，也可能成为可行的消费业务。

Their key design insight is to replace rigid narrative plans with hooks, character rules, memory, and room for improvisation. Peper puts it plainly: “I am not the writer. The Tolan is the writer and the actor.” The human creator's job becomes teaching the character how to tell the best story in each moment. Peper also finds AI most useful around human creation, as an endlessly patient brainstorming partner before writing and a copy editor afterward, rather than as the author of the final scene.

他们最关键的设计洞见，是用钩子、角色规则、记忆和即兴空间取代僵硬的叙事计划。Peper 直言：“I am not the writer. The Tolan is the writer and the actor.” 人类创作者的工作，转变为教会角色如何在每个当下讲出最好的故事。Peper 还发现，AI 最适合围绕人类创作发挥作用：写作前充当耐心无限的头脑风暴伙伴，写作后充当 copy editor，而非直接成为最终场景的作者。

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

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