八月三日

Vercel CEO Guillermo Rauch says every day-to-day job at the company now involves its internal agent, 𝚟, with daily interactions and token use growing exponentially. The system spans finance, communications, documentation, marketing, engineering, and business analytics while maintaining per-user memories, workflows, and schedules. The more consequential design choice is consolidation: after teams created dozens of separate agents, Vercel made 𝚟 both the primary interface and a router that delegates to sub-agents and skills. Rauch compares the pattern to a monolith or a company's main domain, with purpose-built agents reserved for the equivalent of subdomains. He argues that if agents become foundational to companies, ownership of the full path from source and runtime to data and tokens will matter more than adopting someone else's generic integration.

Vercel CEO Guillermo Rauch 表示，公司如今每一项日常工作都会用到内部 agent 𝚟，其每日交互量和 token 使用量都在指数增长。这个系统覆盖财务、沟通、文档、营销、工程和商业分析，同时为每位用户保留独立的记忆、工作流与日程。更重要的设计选择是统一入口：在各团队陆续构建出几十个独立 agent 后，Vercel 让 𝚟 同时充当主要交互界面和 router，再把任务委派给 sub-agent 与 skill。Rauch 把这种模式类比为 monolith 或企业的主域名，只有少数专用 agent 才值得拥有类似子域名的独立入口。他认为，如果 agent 最终成为公司的基础，那么从 source、runtime 到 data、token 的全链路控制权，会比接入别人的通用 integration 更重要。

Box CEO Aaron Levie argues that math, cybersecurity, and coding may be automated earlier precisely because their outputs can be tested objectively. Verifiability supplies clearer reward signals during training and scalable checks during deployment. Legal negotiation, marketing, sales, and budgeting are harder in a different way: they depend on changing context, operator judgment, risk tolerance, and outcomes that may not become observable for months. Levie's implication is that stronger models alone will not unlock those domains; the applied AI layer must redesign processes and invent ways to evaluate knowledge work over time. Anthropic's Thariq sees the complementary economic effect already emerging in mathematics: as AI makes more work possible and easier to understand, Jevons paradox could raise rather than reduce demand for people who can reason about math at higher levels.

Box CEO Aaron Levie 认为，数学、cybersecurity 和 coding 反而可能因为结果可以被客观验证，而比其他工作更早实现自动化。可验证性既能在训练时提供更清晰的 reward signal，也能在部署时支持可扩展的结果检查。法律谈判、营销、销售和预算制定则以另一种方式更难：它们依赖持续变化的上下文、操作者判断、风险偏好，而且结果可能几个月后才显现。Levie 的结论是，仅靠更强模型无法解锁这些领域；applied AI layer 还需要重构流程，并发明能够长期评估知识工作的机制。Anthropic 的 Thariq 则看到这一逻辑在数学领域产生了互补的经济效应：当 AI 让更多数学工作得以开展、也更容易理解时，Jevons paradox 可能让能够进行高层次数学思考的人才需求不降反升。

Every CEO Dan Shipper describes a three-stage response when a model suddenly performs work that once required constant human involvement. First comes an "agency rupture," when the model's achievement dominates attention and the surrounding human contribution disappears. Next, users recognize the scaffolding needed to direct the model and obtain reliable quality. Finally, they reconstruct agency around judgment, orchestration, and a higher standard of output, until AI use becomes implicit in saying "I did this." Shipper's practical thesis is that the ability to turn repeated capability shocks into curiosity and play is a leading indicator for thriving in the AI economy. He also expects changes in technological capability to reshape moral intuitions about what people ought to do, not only what they can do.

Every CEO Dan Shipper 把模型突然接管原本需要人类全程参与的工作后，人们的反应概括为三个阶段。第一阶段是 “agency rupture”：模型的成就占据全部注意力，围绕它的人类贡献暂时变得不可见。第二阶段，用户开始看见引导模型、保证质量所需的 scaffolding。第三阶段，人们围绕判断、orchestration 和更高的质量标准重建主体性，直到使用 AI 已经隐含在一句 “这是我做的” 之中。Shipper 的实际判断是，能否把反复出现的能力冲击转化为好奇心和玩心，是一个人能否在 AI 经济中成长的先行指标。他还认为，技术不仅改变人能做什么，也会随之改变人们对于“应该做什么”的道德直觉。

FPV Ventures partner Nikunj Kothari says early- and mid-stage venture funding has become "vibes capital," with fundraising outcomes increasingly detached from company fundamentals and heavily dependent on whether a startup occupies an in-fashion sector. He expects the distortion to persist for at least 12 to 18 months because dry powder and the AI tailwind keep capital abundant and competition intense. Public markets show a parallel acceleration: even trillion-dollar companies can swing more than 5% on sentiment and model releases, while sector rotations get faster. His long-run advice remains conventional but more urgent in the short run: profitability, a clean cap table, and control over one's destiny still win eventually, yet founders must understand that capital is currently being used as a competitive weapon.

FPV Ventures partner Nikunj Kothari 表示，早期与成长期 VC 已经变成了 “vibes capital”：融资结果越来越脱离公司基本面，更取决于 startup 是否处于当下热门赛道。他预计这种扭曲至少还会持续 12 到 18 个月，因为充足的 dry powder 与 AI 顺风会继续推高资本供给和竞争强度。公开市场也在同步加速：即使是万亿美元公司，也可能因情绪和模型发布出现超过 5% 的波动，板块轮动则越来越快。他的长期建议依旧传统，但在短期内更为迫切：盈利能力、干净的 cap table 和对自身命运的控制最终仍会胜出，不过 founder 必须清楚，当前资本正被当作竞争武器使用。

Peter Yang highlights Hermes, an open-source agent that creates its own skills, and the background Hermes Curator that periodically inspects those skills and memories for clutter and inefficiency. Because users can supply their own definition of low-quality material, the system can rewrite its cleanup loop around their preferences. This turns memory maintenance from a manual housekeeping task into a customizable part of the agent itself, addressing the degradation that can follow when a long-running system keeps accumulating instructions and artifacts.

Peter Yang 介绍了能够自行创建 skill 的开源 agent Hermes，以及在后台定期检查 skill 与 memory 中冗余和低效内容的 Hermes Curator。由于用户可以提供自己对低质量内容的定义，这个系统还能按照个人偏好重写清理循环。它把 memory 维护从人工整理任务变成 agent 自身可定制的一部分，直接处理长期运行的系统不断积累指令与 artifact 后可能发生的质量退化。

Swyx reports using Codex computer use to handle a support chat and push for faster resolution. The agent responded to an attempt to assign fault by presenting the relevant receipts, while the support staff apparently did not realize they were interacting with a bot. The example is small but revealing: browser agents are moving beyond scripted form filling into live negotiation, where they must preserve evidence, respond to resistance, and pursue an outcome across a human conversation.

Swyx 分享了一个 Codex computer use 实例：agent 代他处理客服聊天并争取更快解决问题。当客服试图把责任归到用户一方时，agent 用完整凭据进行了回应，而对方似乎没有意识到自己正在与 bot 对话。这个案例虽小，却很能说明趋势：browser agent 正从按脚本填写表单，走向实时协商场景，需要保存证据、应对阻力，并在与人类的连续对话中主动追求结果。

Replit CEO Amjad Masad put his LLM chess engine on LiChess, where it autonomously plays real games against both people and bots. At the time of the post it had reached a 1253 Elo rating and was capable of playing three games concurrently. The project offers a concrete, externally scored test of an agent's ability to maintain state, choose actions under time pressure, and operate continuously in a live environment.

Replit CEO Amjad Masad 已经把自己的 LLM chess engine 接入 LiChess，让它自主与真人和 bot 进行真实对局。发帖时，这个系统达到 1253 Elo，并能够同时进行三盘棋。这个项目提供了一项具体且由外部计分的 agent 测试，用来观察系统能否持续维护状态、在时间压力下选择行动，并在 live environment 中连续运行。

The Takeaway: Core Automation founders Jerry Tworek and Rohan Anil believe the next model frontier requires architectural research that enables systems to learn from users and real-world tasks at test time, and they are building an automation-native lab to search for it faster.

Tworek, formerly a VP at OpenAI who worked on reasoning research, says the industry has mastered large-scale pretraining and reinforcement learning, but architecture is now the bottleneck. His change of mind came from watching benchmark scores rise without real-world tasks being solved at the same rate. Training and evaluations covered similar distributions, while deployed work remained messier. His conclusion is direct: "We need to have models that learn at test time." In-context learning is data-efficient but limited by context size; continuous fine-tuning is inefficient and risks catastrophic forgetting. He therefore wants architectures capable of longer-horizon adaptation.

Anil, a former Gemini pretraining lead with experience at Google Brain and Anthropic, frames architecture as a decision about how computation is spent. Autoregressive chain-of-thought adds compute one token at a time, while current RL can consume long rollouts to deliver very little feedback. He sees room for large efficiency gains by jointly redesigning architecture and optimization, but insists that any theoretically superior design must also run efficiently on real hardware.

Core Automation's lab is built around that constraint. The team wants to rebuild the deep-learning stack, automate experiment execution, and increase iteration from roughly one architectural experiment a day toward ten or even one hundred. Its nearer-term target is kernel generation: Anil describes a QR-kernel effort where human-guided search reached seven times baseline performance, but roughly $100,000 of coding-agent work over four weeks, guided by rare expert judgment, was needed to reach a 60-times result. For Tworek, genuine AGI begins when a model can improve itself without humans in the loop; today's human-LLM hybrid is productive, but autonomous improvement remains far away.

核心结论： Core Automation founders Jerry Tworek 与 Rohan Anil 认为，下一阶段的模型突破需要新的架构，让系统能够在 test time 从用户与真实任务中持续学习；他们正在打造一家 automation-native lab，以更快速度搜索这条路径。

Tworek 曾任 OpenAI VP，并参与 reasoning research。他认为行业已经掌握大规模 pretraining 与 reinforcement learning，但“通往更强模型与更智能系统的瓶颈，就是架构本身”。他之所以改变判断，是因为亲眼看到 benchmark 分数持续上升，真实世界任务却没有以同样速度被解决。训练集与 eval 覆盖的是相似分布，而实际部署中的工作更加混乱。In-context learning 的数据效率很高，却受 context size 限制；continuous fine-tuning 则效率低下，并有 catastrophic forgetting 风险。因此，他希望找到能够进行长期适应的新架构。

Anil 曾负责 Gemini pretraining，也有 Google Brain 和 Anthropic 经历。他把 architecture 看作“计算资源应当如何花费”的选择。Autoregressive chain-of-thought 逐 token 增加计算，而当前 RL 可能执行很长的 rollout，最终却只向网络传回极少反馈。他认为，把 architecture 与 optimization 联合重构，存在获得巨大效率提升的空间，但任何理论上更优的设计也必须能在真实硬件上高效运行。

Core Automation 的实验室正围绕这个约束搭建。团队希望重建 deep-learning stack、自动执行实验，并把 architecture experiment 的迭代速度从大约每天一次提升到十次，甚至一百次。近期目标之一是 kernel generation：Anil 描述了一项 QR kernel 实验，human-guided search 能做到 baseline 的 7 倍性能，但要达到 60 倍，则需要极少数专家提供判断，并让 coding agent 在四周内消耗约 10 万美元。对 Tworek 而言，真正的 AGI 始于模型能够在没有人类参与的情况下改进自身；今天的 human-LLM hybrid 已经很高效，但 autonomous improvement 仍然非常遥远。