AI Builders Digest — 2026-06-01
X / TWITTER
Thibault Sottiaux, Codex & ChatGPT at OpenAI
Thibault Sottiaux says Codex is crossing a usage milestone: “Five million users would agree,” and OpenAI is resetting limits to celebrate, pushing users to “go /fast.” More importantly, he frames the GPT-5.x sequence as a simple product contract: each increment should improve both capability and token efficiency, which users experience as speed. He also openly asks what Codex has failed to fix for too long, a useful signal that Codex is now in a rapid product-hardening phase, not just a model demo phase.
https://x.com/thsottiaux/status/2060964284117782996
https://x.com/thsottiaux/status/2060960564676034726
https://x.com/thsottiaux/status/2060627747760984429
Thibault Sottiaux 表示 Codex 正在跨过一个使用量里程碑：“Five million users would agree”，OpenAI 也会重置额度来庆祝，并鼓励用户 “go /fast”。更关键的是，他把 GPT-5.x 的版本演进定义成一个清晰的产品承诺：每次小版本递增都应该带来能力提升和 token efficiency 提升，用户感知到的就是更快。他还公开询问 Codex 哪些长期没修的问题最烦人，这说明 Codex 已经进入快速产品打磨阶段，而不只是模型能力展示阶段。
https://x.com/thsottiaux/status/2060964284117782996
https://x.com/thsottiaux/status/2060960564676034726
https://x.com/thsottiaux/status/2060627747760984429
Guillermo Rauch, Vercel CEO
Vercel CEO Guillermo Rauch cuts through the AI hype with a product-first rule: “Ship the best product. Use lots of AI, some AI, maybe no AI. Just be the best.” He also points to per-API-key spend caps on Vercel AI Gateway, which is exactly the kind of unglamorous control layer teams need as AI usage moves from experiments into production budgets.
https://x.com/rauchg/status/2060803480823193840
https://x.com/rauchg/status/2060787704166776927
Vercel CEO Guillermo Rauch 给出了一个反 hype 的产品原则：“Ship the best product. Use lots of AI, some AI, maybe no AI. Just be the best.” 他还提到 Vercel AI Gateway 支持按 API key 设置 spend cap，这类不性感但关键的控制层，正是团队把 AI 从实验推向生产预算时真正需要的基础设施。
https://x.com/rauchg/status/2060803480823193840
https://x.com/rauchg/status/2060787704166776927
Aaron Levie, Box CEO
Box CEO Aaron Levie argues that AI is not only a cost-cutting story inside large enterprises. In the CIO, CTO, and CEO conversations he is seeing, companies are often growing new functions around AI, including FDEs and engineering, or reinvesting efficiency gains into sales, marketing, customer success, risk prevention, and other constrained areas. His warning: companies that only harvest savings may lose to companies that use AI to serve customers better.
https://x.com/levie/status/2060923684295221390
Box CEO Aaron Levie 认为，AI 在大企业里不只是降本故事。他看到的 CIO、CTO、CEO 对话里，很多公司正在围绕 AI 扩张新岗位，比如 FDE 和工程，也会把效率收益重新投入销售、市场、客户成功、风险防控等过去受成本约束的环节。他的提醒是：只把 AI 当省钱工具的公司，可能会输给那些用 AI 更好服务客户的公司。
https://x.com/levie/status/2060923684295221390
Ryo Lu, Design at Cursor
Cursor designer Ryo Lu highlights a small but important product detail in auto-review: Cursor explains the command and the risk, which helps new coders understand what is happening instead of blindly approving agent actions. This is the right direction for AI coding UX: not just “do it for me,” but “make me capable while doing it.”
https://x.com/ryolu_/status/2060766674203353190
Cursor 设计师 Ryo Lu 强调了 auto-review 里的一个小但重要的产品细节：Cursor 会解释命令和风险，让新手 coder 能理解发生了什么，而不是盲目批准 agent 操作。这是 AI coding UX 的正确方向：不只是“替我做”，而是在执行过程中“让我变得更有能力”。
https://x.com/ryolu_/status/2060766674203353190
Peter Steinberger, OpenClaw and OpenAI builder
Peter Steinberger says GPT-5.5, /goal, autoreview, and crabbox have changed his agent workflows from 30 to 60 minute prompts into 4 to 10 hour tasks with much higher confidence in readiness. His practical lesson is blunt: “Yielding agents is a skill.” He also describes a useful debugging pattern: asking Codex to review code may produce “all good,” but telling it there is a bug can make it loop until it finds real issues.
https://x.com/steipete/status/2060678430031597696
https://x.com/steipete/status/2060672154727825718
https://x.com/steipete/status/2060691552486175041
Peter Steinberger 表示，GPT-5.5、/goal、autoreview 和 crabbox 已经把他的 agent 工作流从 30 到 60 分钟的 prompt，推进到经常 4 到 10 小时的长任务，而且他对任务完成质量的信心更高。他的经验很直接：“Yielding agents is a skill.” 他还分享了一个实用 debugging 模式：让 Codex review 代码，它可能说没问题；但如果明确告诉它“这里有 bug”，它会持续循环并更容易找出真实问题。
https://x.com/steipete/status/2060678430031597696
https://x.com/steipete/status/2060672154727825718
https://x.com/steipete/status/2060691552486175041
Dan Shipper, Every CEO
Every CEO Dan Shipper shows how far daily AI coding usage has scaled for power users: 38B tokens, a 56-hour longest task, and a 41-day Codex streak. The signal is less about the vanity metric and more about workflow endurance: frontier coding agents are becoming something builders run continuously, not occasionally.
https://x.com/danshipper/status/2060771279280513362
Every CEO Dan Shipper 展示了高强度 AI coding 用户的使用规模：38B tokens、最长 56 小时任务、连续 41 天 Codex streak。这里的信号不只是炫耀数字，而是工作流耐力：前沿 coding agent 正在从偶尔使用的工具，变成 builders 持续运行的工作系统。
https://x.com/danshipper/status/2060771279280513362
Peter Yang, Product at Roblox
Peter Yang sketches a useful education product intuition: the “ultimate education app” may look like playing Final Fantasy while learning math and CS at the same time. He also points to real motivation dynamics from learning CS with Brilliant, where progress, challenge, and social comparison kept his daughter engaged. The builder takeaway: AI education products may win by becoming games with real curriculum underneath, not by adding quizzes to chatbots.
https://x.com/petergyang/status/2060930599565811774
https://x.com/petergyang/status/2060928818383355907
Peter Yang 提出了一个有价值的教育产品直觉：终极教育 app 可能像玩 Final Fantasy，但同时在学习数学和 CS。他还分享了和女儿用 Brilliant 学 CS 的体验，真正驱动持续学习的是进度、挑战和社交比较。给 builder 的启发是：AI 教育产品可能不是给 chatbot 加测验，而是把真实课程藏在真正好玩的游戏体验下面。
https://x.com/petergyang/status/2060930599565811774
https://x.com/petergyang/status/2060928818383355907
PODCASTS
Unsupervised Learning — Ep 87: Gemini Co-Lead on World Models, RL's Next Domains & Continual Learning
The Takeaway: Oriol Vinyals sees the next frontier less as “more chatbot polish” and more as models that learn from experience, use memory like a file system, and eventually generate their own scaffolding for tasks.
Oriol Vinyals, Gemini co-lead alongside Noam Shazeer and Jeff Dean, frames world models as more than video generation. A useful world model should let language control a simulation of the world, then become valuable for planning, robotics, and prediction before action. But he is careful about the gap: for robotics, the model needs physical precision around grasping, force, and transfer that current systems do not fully have.
His strongest builder advice is around evals and data. For startups deciding whether to work at the model layer, he says the durable value is often not training a model first, but building the evaluations and domain data that define progress. He also argues that memory may evolve through nonparametric knowledge bases, files, folders, retrieval, and skills, rather than every user getting personalized model weights. The quote worth underlining: “We do have access because it’s an agent to a memory system, which is the computer itself.”
On agents, Vinyals expects today’s hand-coded scaffolds, multi-agent systems, delegation patterns, and long-running workflows to become more automated. In the limit, the model may write the scaffold it needs on the fly. That is a sharp signal for builders: the moat may shift from manually designing clever scaffolds to owning the task distribution, evals, memory, and product context where the agent learns to operate.
https://www.youtube.com/watch?v=NQczevdpxq0
The Takeaway：Oriol Vinyals 认为，下一阶段的前沿不只是“把 chatbot 打磨得更好”，而是模型能从经验中学习、像使用文件系统一样使用记忆，并最终为不同任务自动生成自己的 scaffold。
Oriol Vinyals 是 Gemini co-lead，与 Noam Shazeer、Jeff Dean 一起负责 Gemini。他把 world model 定义为不只是视频生成。一个真正有用的 world model，应该让语言能控制对世界的模拟，并进一步用于规划、机器人和行动前预测。但他也很谨慎：在机器人领域，模型还需要对抓取、力、迁移等物理细节有足够精度，目前系统还没有完全做到。
他给 builder 最重要的建议集中在 eval 和 data。对于纠结要不要做 model layer 的创业公司，他认为长期价值往往不是一开始就训练模型，而是建立能定义进展的评估体系和领域数据。他还认为 memory 可能更多通过非参数知识库、文件、文件夹、检索和 skills 演进，而不是每个用户都有一套个性化模型权重。最值得划线的一句话是：“We do have access because it’s an agent to a memory system, which is the computer itself.”
在 agent 方向，Vinyals 预计今天手写的 scaffold、多 agent 系统、delegation 模式和长任务工作流会越来越自动化。极限状态下，模型可能会按需即时写出自己需要的 scaffold。这对 builder 是一个强信号：护城河可能会从手工设计 clever scaffold，转向拥有任务分布、eval、memory 和产品上下文，让 agent 能在其中持续学习和执行。
https://www.youtube.com/watch?v=NQczevdpxq0
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders