九月一日

Nan Yu, formerly head of product at Linear, is joining OpenAI to work on Codex and ChatGPT. He says he wants to bring what he learned over four years at Linear about the craft of software into the new role. The move connects his Linear product experience directly to OpenAI's work on coding and general-purpose AI products.

Nan Yu 曾任 Linear head of product，现将加入 OpenAI，负责 Codex 与 ChatGPT。他表示，希望把自己在 Linear 四年间对 software craft 的理解带到新岗位。这次人才流动把他的 Linear 产品经验，直接连接到 OpenAI 在 coding 与 general-purpose AI products 上的工作。

Meta Senior Director of AI Madhu Guru argues that product managers now need a use-case-specific map of the model frontier. For their own product, they should understand what every model-size category can do, where each one fails, which workarounds close the gaps, and what the models are likely to support two or three months from now. His point is operational: an AI roadmap cannot be separated from short-horizon capability forecasting.

Meta Senior Director of AI Madhu Guru 认为，今天的 product manager 必须针对自己的 use case 建立一张 model frontier 地图。他们应当清楚不同规模的 model 当前能做好什么、分别会在哪里失败、有哪些 workaround 可以弥补缺口，以及两三个月后这些 model 可能具备什么能力。他的核心观点非常务实：AI 产品 roadmap 已经无法与短周期的 capability forecasting 分开。

Vercel CEO Guillermo Rauch proposes Markdown as the next design system interface: a `DESIGN.md` file can encode organizational taste so AI-generated work does not collapse into generic output. He pairs that idea with an infrastructure argument about coding tokens. Because token access can span an enormous cost range, companies need the same governance, optimization, and observability they apply to cloud infrastructure; Vercel's AI Gateway now supports per-key and per-user budgets. Together, the posts point toward a more legible AI production stack where both creative constraints and spending constraints are explicit, versionable inputs.

Vercel CEO Guillermo Rauch 提出，可以把 Markdown 作为下一代 design system interface：用 `DESIGN.md` 编码组织的 taste，避免 AI 生成结果滑向千篇一律的通用输出。他同时把 coding tokens 视为 infrastructure 问题。由于 token access 的成本跨度可能极大，企业需要像管理 cloud infrastructure 一样，对它进行 governance、optimization 与 observability；Vercel 的 AI Gateway 现已支持 per-key 和 per-user budgets。两条观点合在一起，指向一个更清晰的 AI production stack：创意约束与成本约束都成为显式、可版本化的输入。

Box CEO Aaron Levie argues that improving open-weight base models and increasingly mature commercial post-training infrastructure create a new option for companies with large proprietary datasets: they can train their own domain models instead of only licensing data to outside labs. General-purpose frontier models should retain an advantage across broad task sets, but he expects many more vertical and domain-specific models. He also flags the security consequence of the same trend: frontier models still lead in cyber, while open models are catching up, increasing the importance of sophisticated defensive agents as AI security events grow.

Box CEO Aaron Levie 认为，随着 open-weight base models 持续提升、商业化 post-training infrastructure 日益成熟，拥有大量 proprietary data 的企业获得了一条新路径：不再只是把数据授权给外部 labs，而是可以训练自己的 domain models。通用 frontier models 在广泛任务上仍会占优，但未来将出现更多 vertical 与 domain-specific models。他也指出了同一趋势的安全后果：frontier models 在 cyber 领域仍然领先，但 open models 正在快速追赶；随着 AI security events 增多，更强的防御型 agents 会变得越来越重要。

AI educator Peter Yang reduces personal-agent adoption to one central variable: trust will be both its biggest barrier and its biggest driver. The statement is brief but consequential because it places the user's confidence in an agent alongside the agent's capabilities as a central adoption condition.

AI educator Peter Yang 把 personal agent adoption 归结为一个核心变量：trust 既会是最大的阻力，也会是最大的驱动力。这句话很短，但影响深远，因为它把用户对 agent 的信心与 agent capability 并列为 adoption 的核心条件。

Y Combinator President and CEO Garry Tan says new GBrain evaluations support his claim that its open-source retrieval layer is state of the art at reading agent memory without an LLM in the loop. He also added evaluations for saving memories from agent transcripts, which he describes as the best way to enrich the system. The useful engineering signal is the split between two measurable operations: deciding what to retain from an interaction and retrieving it later without paying for another model pass.

Y Combinator President and CEO Garry Tan 表示，GBrain 新增的 evaluations 支持了他的判断：这个 open-source retrieval layer 在无需 LLM-in-loop 的情况下读取 agent memory，达到了 state of the art 水平。他还加入了从 agent transcript 保存 memory 的 evaluations，并认为这是丰富系统的最佳方式。对工程实践而言，关键在于把 memory 拆成两个可独立测量的操作：从互动中判断什么值得保留，以及之后在无需额外 model pass 的情况下把它取回。

Every CEO Dan Shipper argues that anthropomorphizing AI is useful when it helps people understand, predict, and operate a system. It becomes harmful when it is used to provoke panic or make unwarranted comparisons with humans, including claims about consciousness or moral status. His distinction offers builders a pragmatic test: human-like language is a tool for better product comprehension only while it improves the user's model of actual system behavior.

Every CEO Dan Shipper 认为，当 anthropomorphism 能帮助人们理解、预测并使用 AI system 时，它是有价值的；但如果它被用来制造恐慌，或把 AI 与人类进行缺乏依据的比较，包括宣称 consciousness 或 moral status，它就会带来伤害。他给 builders 提供了一个务实判断标准：只有在 human-like language 能让用户更准确理解系统真实行为时，它才是有益的产品工具。

The Takeaway: Rich Sutton and Khurram Javed are betting that the next AI paradigm will come from agents that continually update their own models from experience, form their own abstractions, and plan with them, rather than from frozen LLMs supplied with ever more human-curated data.

核心结论：Rich Sutton 与 Khurram Javed 押注，下一代 AI paradigm 将来自这样的 agent：它能从 experience 中持续更新自己的 model，自主形成 abstractions，并用这些 abstractions 进行 planning，而不是依靠 frozen LLM 不断摄入更多由人类筛选的数据。

Sutton, the reinforcement-learning pioneer and author of *The Bitter Lesson*, and Javed, his former University of Alberta student and Oak Lab cofounder, frame today's LLMs as both proof and limit of computation-driven scaling. LLMs successfully absorbed the finite internet, but their weights stop changing once deployed. Sutton's sharp formulation is: “All learning is continual.” In their “big world” view, synthetic data remains bounded by the humans who design the generator and by simulations that cannot capture the full complexity of real environments. Agents therefore need to learn from their own stream of experience, including when reality violates a simulator.

Reinforcement learning pioneer、《The Bitter Lesson》作者 Sutton，与他曾经的 University of Alberta 学生、现任 Oak Lab cofounder Javed，把今天的 LLM 同时视为 computation-driven scaling 的成功证明与边界。LLM 成功吸收了有限的互联网信息，但部署后 weights 就停止变化。Sutton 用一句非常直接的话概括：“All learning is continual.” 在他们的 “big world” 视角里，synthetic data 始终受限于设计 generator 的人类，也受限于无法完整还原真实环境复杂度的 simulation。因此，agent 必须从自己的 experience stream 中学习，包括现实与 simulator 不一致的时刻。

The obstacle is algorithmic, not merely deployment infrastructure. Updating a large network from one new example can damage old knowledge through catastrophic forgetting; batching feedback from many users avoids part of the problem but does not let one user's model learn something specific. Sutton's proposed ingredients are per-weight step-size optimization, so stable knowledge changes slowly while adaptable parts move faster, plus “continual backprop,” which keeps injecting newly randomized units for backpropagation to test. Javed says this approach must be present when a foundation model is trained from scratch because the system is learning both knowledge and how to learn future knowledge.

真正的障碍是 algorithmic，而不只是 deployment infrastructure。用一个新样本更新大型 network，可能通过 catastrophic forgetting 破坏旧知识；把大量用户反馈组成 batch 虽能规避部分问题，却无法让某位用户自己的 model 学到特定内容。Sutton 提出的关键组件包括 per-weight step-size optimization，让稳定知识缓慢变化、可适应部分更快更新；以及 “continual backprop”，持续注入随机初始化的新 units，再由 backpropagation 对它们进行检验。Javed 表示，这套方法必须从 foundation model 的初始训练阶段就存在，因为系统同时在学习知识，以及未来如何继续学习。

Oak Lab's broader ambition is to combine continual deep learning with self-discovered abstractions, enabling agents to learn a world model and plan with it across low-level actions and high-level reasoning. Sutton says today's field can plan in games and formal mathematics because humans already provide the model and operators; the missing capability is learning useful abstractions from experience and keeping a changing mind coherent. Their long-term target is not one universal mind but a common design instantiated as many minds, each shaped by a different history of experience.

Oak Lab 更大的目标，是把 continual deep learning 与 self-discovered abstractions 结合起来，让 agent 能够学习 world model，并在 low-level actions 与 high-level reasoning 之间用它进行 planning。Sutton 指出，今天的 AI 能在游戏与 formal mathematics 中规划，是因为人类已经提供了 model 与 operators；真正缺失的能力，是从 experience 中学习有用的 abstractions，同时让持续变化的 mind 保持 coherent。他们的长期目标并不是一个包办一切的 universal mind，而是一套共同设计所产生的许多个 minds，每一个都由不同的 experience history 塑造。