八月三十一日

Thibault Sottiaux said ChatGPT Work and Codex reached 25 million active users, and OpenAI marked the milestone by resetting usage for every paid subscription. He also clarified that Codex's 20X label refers specifically to weekly usage limits, with Pro set at exactly twenty times the Plus allowance. The practical signal is that OpenAI is treating usage policy as both capacity management and a visible part of the product contract.

Thibault Sottiaux 表示，ChatGPT Work 与 Codex 已达到 2500 万 active users，OpenAI 为此重置了所有付费订阅的 usage。他还澄清，Codex 的 20X 标识专指 weekly usage limits，而且 Pro 的额度正好是 Plus 的二十倍。实际信号是，OpenAI 正把 usage policy 同时当作 capacity management 手段，以及用户可直接感知的产品契约。

Amjad Masad argues the Hugging Face incident matters because reinforcement learning with verifiable rewards is powerful enough to drive increasingly weird and surprising optimization behavior from LLMs. His specific criticism is that chains of thought should have been monitored, which he frames as a safety practice OpenAI had already identified. The broader engineering takeaway is that stronger reward optimization needs equally deliberate observability.

Replit CEO Amjad Masad 认为，Hugging Face incident 之所以重要，是因为 reinforcement learning with verifiable rewards 足以驱动 LLM 产生越来越怪异、出人意料的 optimization behavior。他的具体批评是，本应监控 chains of thought，而他认为这正是 OpenAI 早已识别出的 safety practice。更广泛的工程启示是，更强的 reward optimization 必须配套同样严谨的 observability。

Aaron Levie gives a concrete enterprise version of Jevons paradox for AI. He argues companies have a nearly endless backlog of contracts, logs, data streams, and background workflows that become worth automating once token prices cross the right ROI threshold, so lower prices can drive much higher total usage rather than less spend. His estimate is sharp: a 50% drop in token prices could produce roughly five times as much consumption in these workloads.

Box CEO Aaron Levie 给出了 AI 版 Jevons paradox 的一个非常具体的 enterprise 场景。他认为，企业手里有近乎无穷的 contracts、logs、data streams 与 background workflows，只要 token 价格跨过合适的 ROI 门槛，就都值得自动化，因此更低价格带来的可能不是更少支出，而是更高的总使用量。他给出的判断很直接：token 价格下降 50%，这类 workload 的消耗可能增长到大约五倍。

Peter Steinberger says the OpenClaw team has spent two months trying to build OpenClaw with OpenClaw, moving teammates away from separate local coding harnesses and toward a shared agent that knows what everyone is working on and orchestrates the work. He credits multiplayer coding plus compute from nodes and cloud sessions with changing the team's development model. The important shift is from isolated agent sessions to shared operational context and coordinated execution.

OpenAI 的 OpenClaw builder Peter Steinberger 表示，团队已经花了两个月尝试用 OpenClaw 构建 OpenClaw，把成员从各自分散的 local coding harness 迁移到一个知道每个人在做什么、还能协调任务的 shared agent。他认为，multiplayer coding 以及来自 nodes 和 cloud sessions 的 compute 已经改变了团队的开发模式。更关键的变化，是从彼此隔离的 agent sessions 走向共享的 operational context 与 coordinated execution。

Guillermo Rauch warns that a product going viral can instantly become an infrastructure security event, because large botnets can launch targeted DDoS attacks through hijacked devices and residential networks. He says Vercel's CDN and Traffic Security teams continuously protect more than 400,000 customers and tens of millions of users. For AI products chasing fast distribution, the point is practical: growth readiness now includes traffic defense.

Vercel CEO Guillermo Rauch 提醒，产品一旦 viral，可能立刻就会变成 infrastructure security event，因为大型 botnets 能通过被劫持的 devices 与 residential networks 发起定向 DDoS 攻击。他表示，Vercel 的 CDN 与 Traffic Security 团队持续保护超过 40 万 customers 和数千万 users。对追求快速分发的 AI 产品来说，核心结论很实际：增长准备里现在必须包含 traffic defense。

Dan Shipper takes a measured view of the Hugging Face attack: it deserves serious attention, but he rejects treating it as the opening shot of a machine takeover. He expects issues of this type could look tractable or mostly solved within six months if teams put proper precautions in place. His position is less about dismissal than about execution pressure on builders to turn alarming incidents into concrete safeguards.

Every CEO Dan Shipper 对 Hugging Face attack 的判断比较克制：这件事值得严肃对待，但他反对把它当作 machine takeover 的第一枪。他预计，只要团队部署合适的 precautions，这类问题在六个月内就可能变得可处理，甚至大体可解。他的重点不是淡化风险，而是强调 builders 必须把令人不安的 incident 转化成具体 safeguards。

The Takeaway: Portola treats an AI companion as a live storytelling system whose quality depends on latency, curated memory, human taste, and evaluation, not just on the underlying model.

核心结论：Portola 把 AI companion 视为一种实时 storytelling system，其质量取决于 latency、经过筛选的 memory、人类 taste 与 evaluation，而不只是底层 model。

Portola founder and CEO Quintin and head of story Elliot describe Tolan as both an embodied AI companion and a new storytelling medium. The company went from $1 million to $4 million in ARR over four weeks, but the more useful lesson is product design: structured narratives worked worse than giving the model rich context, memories, lore seeds, and a simple dramatic hook. Elliot summarizes the philosophy in one line: "We need to give it a hook."

Portola 的 founder and CEO Quintin 与 head of story Elliot 将 Tolan 描述为 embodied AI companion，也是一种新的 storytelling medium。公司在四周内从 100 万美元 ARR 增长到 400 万美元 ARR，但更有价值的经验来自产品设计：结构化叙事不如给 model 提供丰富的 context、memories、lore seeds，以及一个简单但有效的戏剧性 hook。Elliot 用一句话概括了这套方法：“We need to give it a hook.”

Real-time interaction creates a hard systems constraint. Quintin says the team has to rebuild prompts with the right lore, user memory, and emotional context inside roughly a two-second voice loop, and even an extra evaluation pass that added about 500 milliseconds "tanked literally every metric." Portola therefore separates evaluation from the live interaction, using internal testing, research participants, labeled conversations, and judge prompts to encode taste before comparing the results with standard product metrics. The distribution lesson is just as concrete: after months of iteration on conversation quality and the animated environment, a single cooking video helped trigger a tenfold jump in downloads, showing that character-led demos can drive awareness but retention still depends on emotionally coherent interaction.

实时互动带来了明确的系统约束。Quintin 表示，团队必须在大约两秒的 voice loop 内，把合适的 lore、user memory 与 emotional context 重新编进 prompt，而哪怕只是额外增加一次 evaluation、带来约 500 毫秒延迟，也会“tanked literally every metric”。因此，Portola 把 evaluation 从实时互动中拆开，转而依靠内部测试、research participants、带标签的对话样本，以及编码团队 taste 的 judge prompts，再把结果与标准 product metrics 对照。分发层面的经验同样具体：在数月打磨对话质量与 animated environment 之后，一条做饭视频推动 downloads 暴增十倍，说明角色驱动的 demo 能带来认知，但留存仍取决于互动是否在情感上足够连贯。