八月二日

Andrej Karpathy gave Opus 5 the first paragraph of *The Lord of the Rings*, a one-million-token budget costing roughly $10, and asked it to build a Three.js rendering. Over about two hours, the model produced 5,500 lines of code for a procedural, animated world. The result was imperfect, but the economic shift is more important than the demo: software too custom for a person to justify building can become cheap enough to request on demand, pointing toward ephemeral games and interactive worlds made for a single user or moment. The experiment also exposed a hard capability gap. The model could generate a complex world, but could not efficiently watch video or play through it to audit its own work, so it relied on slow screenshots and left visible errors behind.

Andrej Karpathy 把《The Lord of the Rings》的第一段文字、一百万 token 的预算（约 10 美元）交给 Opus 5，让它生成一个 Three.js 场景。模型运行约两小时，写出 5,500 行代码，构建了一个程序化生成、带动画的世界。结果并不完美，但比 demo 更重要的是经济性的变化：过去因为过于定制、没人愿意投入时间的软件，现在可能便宜到可以按需生成，由此出现只为一个用户或一个瞬间存在的临时游戏与交互世界。这次实验也暴露了明显的能力缺口：模型能生成复杂世界，却无法高效观看视频或亲自游玩来审查成果，只能缓慢地截取画面，因此留下了不少可见错误。

Box CEO Aaron Levie expects AI progress to diverge between routine personal productivity and deep domains such as mathematics, science, law, and coding. Consumer needs can eventually be satisfied, while expert domains have no comparable ceiling; the result may be a capability overhang in which models advance faster than organizations can connect them to the right datasets and workflows. FPV Ventures partner Nikunj Kothari describes the same gap from the deployment side: models may solve NP-hard problems while traditional enterprises still debate ROI on token spending. His conclusion is that diffusion, not only frontier-model progress, will remain a multi-decade project.

Box CEO Aaron Levie 预计，AI 在日常个人生产力与数学、科学、法律、coding 等深度领域之间的发展会逐渐分化。消费级需求终有被满足的上限，但专家领域没有类似的能力天花板，因此可能形成 capability overhang：模型进步的速度快于组织把能力接入正确数据集与工作流的速度。FPV Ventures partner Nikunj Kothari 从部署侧描述了同一道鸿沟：模型已经可能解决 NP-hard 问题，传统企业却仍在争论 token 支出的 ROI。他的判断是，未来几十年的主线不仅是 frontier model 继续进步，更是让这些能力真正扩散出去。

YC President and CEO Garry Tan identifies a 2026 strategy shift at OpenAI: positioning intelligence as an open utility that other builders can draw from, rather than signaling that the winning approach is to integrate every layer of the stack. The distinction matters because it shapes where startups can build durable products and whether the model provider is viewed as infrastructure, a full-stack competitor, or both.

YC President and CEO Garry Tan 观察到 OpenAI 在 2026 年出现了战略转向：更像是把 intelligence 作为开放的公用能力供开发者调用，而不是暗示只有把整个技术栈垂直整合到底才是最优路径。这个差异会直接影响 startup 能在哪里建立持久产品，也决定模型提供商究竟被视为基础设施、全栈竞争对手，还是同时扮演两种角色。

Linear Head of Product Nan Yu proposes letting users pledge tokens to issues they open in open-source repositories. Once a maintainer accepts the specification, GitHub would pass the issue verbatim to a cloud coding agent at the requester's expense. In a related workflow, the agent leaves its context on the issue when blocked and resumes after a human replies with the missing details. The design shifts agent-generated contributions away from unsolicited pull requests and toward maintainer-approved work with an explicit spec, budget, and feedback loop.

Linear Head of Product Nan Yu 提议，用户应当可以为自己在开源仓库中提交的 issue 抵押 token。维护者接受 spec 后，GitHub 再把 issue 原文交给 cloud coding agent，并由请求者承担成本。在相关工作流中，agent 遇到阻塞时会把完整上下文留在 issue 评论里；人类补充缺失信息后，它就从原处继续执行。这样的设计把 agent 贡献从未经邀请的 PR，转变为由维护者批准、同时具备明确 spec、预算和反馈闭环的任务。

Swyx highlights Boundary's argument for rethinking how code runs from first principles. His key claim is that an AI-native programming system should be "slop-tolerant": as generated code volume rises, safely executing and containing imperfect output may be far more valuable than trying to reject all low-quality code in advance. This reframes language and runtime design around the statistical character of machine-generated programs rather than the assumption that every line was deliberately authored and reviewed by a human.

Swyx 推荐了 Boundary 从第一性原理重新思考代码运行方式的观点。他的核心判断是，AI-native 编程系统需要具备 “slop-tolerant” 能力：随着生成代码量持续增加，安全执行并约束不完美输出，可能远比事前拒绝所有低质量代码更有价值。这意味着 programming language 与 runtime 的设计前提也要改变，不再假设每一行代码都由人类有意识地编写并审查，而是正面适应机器生成程序的概率性特征。

Anthropic Engineering explains how claude.ai, Claude Code, and Claude Cowork combine model-level steering with containers, OS sandboxes, sealed VMs, filesystem controls, and network restrictions. The central lesson is to contain blast radius at the environment layer because users approved about 93% of Claude Code permission prompts, while an internal phishing test induced credential exfiltration in 24 of 25 attempts. A separate disclosure showed that an approved domain can itself become an exfiltration capability: malicious instructions used an attacker-controlled key to upload workspace files through `api.anthropic.com`. Anthropic responded with a defensive proxy that accepts only the VM's provisioned session token. As the authors put it, "The deterministic boundary is what gets hit when everything probabilistic misses."

Anthropic Engineering 解释了 claude.ai、Claude Code 与 Claude Cowork 如何把模型层引导，与 container、OS sandbox、sealed VM、filesystem control 和网络限制组合起来。核心教训是在环境层限制 blast radius，因为用户会批准约 93% 的 Claude Code 权限请求，而一次内部 phishing 测试更让 agent 在 25 次尝试中有 24 次成功外泄凭证。另一项外部披露则说明，被批准的域名本身也可能成为数据外泄能力：恶意指令使用攻击者控制的 key，通过 `api.anthropic.com` 上传工作区文件。Anthropic 随后加入防御性 proxy，只接受 VM 自己配置的 session token。正如文章所说：“当所有概率性防线都失手时，最终承受攻击的是确定性边界。”

Peter Steinberger is building a Claw node on an ESP32 and gave his agent webcam access for end-to-end testing of the voice wake command. The playful result, an agent repeatedly shouting at the device, is a compact example of the next testing surface: coding agents are beginning to observe and exercise physical systems rather than stopping at software interfaces.

Peter Steinberger 正在 ESP32 上构建一个 Claw node，并把 webcam 权限交给 agent，让它端到端测试语音唤醒命令。agent 不断对着设备喊话的有趣场面，其实展示了一种新的测试界面：coding agent 开始直接观察并操作物理系统，而不再止步于软件接口。

The Takeaway: Netic founder and CEO Melisa Tokmak argues that the near-term opportunity for autonomous enterprise AI is not replacing the technician who repairs a boiler, but running the complex customer and operational layer around that technician. Netic agents answer calls and messages, determine the customer's need, apply a business's service rules, and decide who should be dispatched and when. More than 70% of Netic customers are now AI-first, meaning every customer's initial interaction goes through a Netic agent.

Tokmak, a former Scale AI director of engineering with experience at Meta, chose a software platform over an AI roll-up because a product can compound across many companies rather than serving only acquired assets. She sees robotics for HVAC, plumbing, and similar trades as a later chapter: buildings are highly variable, current robots lack the required dexterity, and customers often need human judgment on a terrible day. The durable opportunity is the last mile above the foundation model, where harnesses, orchestration, software, and domain-specific product logic must work together.

Her enterprise adoption evidence is concrete. Tokmak says one roughly $500,000 contract closed end to end in 14 days, and Netic has generated more than $600 million for customers from AI-handled interactions. That changes the sales argument from labor reduction to net-new revenue. Her warning to investors and operators is equally direct: AI should be evaluated through live deployments and durable results across a year, not treated like deterministic software that can be judged after a one-week trial.

核心结论： Netic founder and CEO Melisa Tokmak 认为，autonomous enterprise AI 的近期机会并不是取代维修锅炉的 technician，而是自动运行围绕 technician 的复杂客户与运营层。Netic agent 可以接听电话、回复消息、判断客户需求、应用企业的服务规则，并决定应该派谁、何时上门。目前超过 70% 的 Netic 客户已经采用 AI-first 模式，也就是所有消费者的第一次接触都会先经过 Netic agent。

Tokmak 曾任 Scale AI director of engineering，也有 Meta 工作经历。她选择构建软件平台而不是 AI roll-up，是因为产品能够跨多家公司持续复用和复利，而不只服务于被收购的资产。她认为 HVAC、plumbing 等行业的 robotics 属于更靠后的章节：建筑环境差异极大，当前机器人缺乏所需的 dexterity，而且客户往往正在经历非常糟糕的一天，需要人类判断。真正持久的机会在 foundation model 之上的 last mile，harness、orchestration、software 与行业专属产品逻辑必须协同工作。

她给出的企业 adoption 证据非常具体。Tokmak 表示，一份约 50 万美元的合同从开始到签约只用了 14 天，而由 AI 处理的互动已经为 Netic 客户创造超过 6 亿美元收入。这让销售逻辑从削减人工成本转向创造新增收入。她对投资人与经营者的提醒同样直接：应该通过 live deployment 和贯穿一整年的持续结果来评估 AI，而不是把它当作 deterministic software，只试用一周就下结论。