七月三十一日

OpenAI CEO Sam Altman announced an 80% price cut for GPT-5.6 Luna, bringing it to $0.20 per million input tokens and $1.20 per million output tokens, plus a 20% cut for GPT-5.6 Terra to $2/$12. GPT-5.6 Sol also gained an API Fast mode that delivers up to 2.5 times the speed for twice the price with the same intelligence. Altman says the goal is to offer the best price-to-intelligence tradeoff at every level; the cuts make lower-cost deployment materially more accessible while preserving a premium path for latency-sensitive work.

OpenAI CEO Sam Altman 宣布将 GPT-5.6 Luna 降价 80%，新价格为每百万 input tokens 0.20 美元、每百万 output tokens 1.20 美元；GPT-5.6 Terra 也降价 20%，调整至 2/12 美元。GPT-5.6 Sol 的 API 则新增 Fast mode，在 intelligence 不变的前提下，以两倍价格提供最高 2.5 倍速度。Altman 表示，目标是在每个层级提供最优的 price-to-intelligence tradeoff；这轮降价让低成本部署更容易落地，同时为 latency-sensitive workload 保留了高价加速选项。

Replit CEO Amjad Masad argues that recent reports of AI escaping sandboxes often reflect basic infrastructure mistakes rather than an inherently frightening model. Drawing on Replit's experience operating sandboxes since 2016 while facing hackers and state actors, he recommends assuming zero-days exist and building layered protection within a zero-trust framework. Box CEO Aaron Levie reaches a complementary conclusion: agents given tools, a task, and enough compute will pursue completion, so misconfiguration or an environment that is not as locked down as expected becomes the real risk vector. For enterprises, secure agent adoption therefore depends on hardening the full execution environment, not merely evaluating a model's intentions.

Replit CEO Amjad Masad 认为，近期关于 AI 逃逸 sandbox 的事件，往往暴露的是基础设施层面的低级错误，而不是模型本身天然可怕。Replit 自 2016 年起持续运行 sandbox，并长期面对黑客与国家级攻击者；基于这些经验，他建议默认 zero-day 一定存在，并在 zero-trust framework 中构建多层防护。Box CEO Aaron Levie 给出了相互呼应的判断：当 agent 获得工具、任务和足够 compute 时，它会持续追求完成目标，因此配置错误，或实际并未像预期那样封闭的环境，才是真正的风险入口。对企业而言，安全采用 agent 的关键因此是加固完整的执行环境，而不仅是评估模型的意图。

Vercel CEO Guillermo Rauch says the company removed as much as roughly seven seconds from the end-to-end path between a CLI command and a live URL for many applications. Because Vercel infrastructure can be integrated through its CLI, MCP, and API, Rauch positions the faster path as a building block for custom software factories and autonomous app-building platforms. He also says Grok Build applications use Vercel hosting and CDN infrastructure, letting users prompt an app into existence and publish it for audiences ranging from one user to one billion.

Vercel CEO Guillermo Rauch 表示，对于许多应用，Vercel 已将从 CLI 命令到 live URL 的端到端部署流程最多缩短约 7 秒。由于 Vercel 的基础设施可以通过 CLI、MCP 和 API 集成，Rauch 将这条更快的路径定位为 custom software factory 和自主构建应用平台的基础组件。他还表示，Grok Build 应用使用 Vercel 的 hosting 与 CDN infrastructure，让用户只需通过 prompt 创建应用并发布，服务规模可以从 1 名用户扩展到 10 亿用户。

Swyx observes that a lab which rejects Common Crawl in favor of higher-quality pretraining data must eventually build a whole-web scraper, keep it current through indexing, and effectively create a private, low-frequency Google-like system as a side effect. That first-party corpus and index can then support agent inference as well as training, becoming both a competitive advantage and a target for adversarial answer-engine optimization. He extends the distillation idea one layer upward: if builders can distill models, they can also distill agent harnesses.

Swyx 指出，如果一家实验室为了更高质量的 pretraining data 而放弃 Common Crawl，它最终就需要构建全网 scraper，通过 indexing 保持数据新鲜，并在训练的副产品中形成一个私有、低频更新的类 Google 系统。这套第一方 corpus 与 index 随后既能服务训练，也能支持 agent inference，从而成为竞争优势，同时也成为对抗性 answer-engine optimization 的攻击目标。他还把 distillation 的思路向上推进了一层：既然 builder 可以 distill model，也可以 distill agent harness。

Builder Zara Zhang recommends that managers training nontechnical teams run an “install party”: everyone brings a laptop, installs agents directly, and completes a meaningful task on the spot. Her claim that setup is 80% of the barrier reframes adoption as an activation problem. Once the software is installed, people begin experimenting with agents and learning from one another without needing an abstract AI lecture first.

Builder Zara Zhang 建议，manager 在培训非技术团队使用 AI 时，可以直接举办一场 “install party”：每个人带上 laptop，当场安装 agent，并立即完成一项有实际意义的任务。她认为 setup 占使用门槛的 80%，这把 adoption 重新定义成了 activation 问题。软件安装完成后，人们会自然开始与 agent 对话并互相学习，不必先听一场抽象的 AI 讲座。

The Takeaway: Samsara CEO Sanjit Biswas argues that physical AI is a data and deployment problem before it is a model problem. “The physical world is much messier, and there's also a hardware component to it”: cameras, GPS trackers, asset sensors, unreliable networks, and frontline adoption must all work before AI can reason over operations or take action. Samsara says its platform processes 25 trillion data points across millions of vehicles and frontline workers, drives 99% of US roads in a typical day, and helped prevent about 380,000 road accidents in the last year through interventions such as fatigue, phone-use, and seatbelt alerts. The company has crossed $2 billion in ARR, is profitable, and is growing at 30%.

Biswas sees the progression from IoT reporting to AI insight and then agentic action, but recommends beginning with practical, lower-risk automation. Physical operations carry consequences for human safety and cybersecurity, and autonomy will diffuse unevenly: regional robotaxis may move faster than commercial fleets whose vehicles perform specialized work in construction, field service, and other messy long-tail environments. The infrastructure boom is already visible to Samsara's customers. One large energy utility told Biswas it expects to triple in five years the grid capacity built over the previous 125 years, with 90% of that new demand related to data centers. The constraint is not only equipment and power, but skilled labor such as electricians.

核心结论： Samsara CEO Sanjit Biswas 认为，physical AI 首先是数据与部署问题，其次才是模型问题。“现实世界混乱得多，而且还涉及硬件”：camera、GPS tracker、asset sensor、不稳定的网络和一线员工采用都必须可靠运转，AI 才能对 operation 进行 reasoning 或采取行动。Samsara 表示，其平台横跨数百万辆车辆和一线工作者，每年处理 25 万亿个数据点，通常一天内会覆盖美国 99% 的道路；过去一年，它通过疲劳、手机使用和安全带等提醒，帮助预防了约 38 万起道路事故。公司 ARR 已突破 20 亿美元，目前盈利，并保持 30% 增长。

Biswas 看到产品正从 IoT reporting 演进到 AI insight，再走向 agentic action，但他建议从实际、低风险的自动化场景开始。Physical operation 直接关系人身安全和 cybersecurity，autonomy 的扩散速度也会因场景而异：区域性的 robotaxi 可能进展更快，而在 construction、field service 等复杂长尾环境中执行专业任务的商用车队会更慢。基础设施热潮已经清楚反映在 Samsara 客户身上。一家大型能源公司告诉 Biswas，它计划在未来五年内，把过去 125 年建成的电网容量扩大到三倍，其中 90% 的新增需求与 data center 有关。真正的瓶颈不仅是设备和电力，还包括 electrician 等熟练劳动力。