Anthropic 发布 MHS 模型硬件标准：让 AI 像插 USB-C 一样操控真实世界设备Anthropic 发布 MHS 模型硬件标准：让 AI 像插 USB-C 一样操控真实世界设备
过去，让 AI 控制一台实验室设备，意味着每个硬件厂商各写一套私有 SDK 与协议。换一家显微镜、换一台移液机械臂，集成就要从零再来——一个自动化流程往往要写数周的胶水代码。MHS 想解决的正是这种碎片化：它定义了一套所有设备都能说、所有智能体都能懂的「共同语言」。
MHS 的核心分三层。其一「驱动（Driver）」：把设备的真实能力——能测什么、能调哪些参数、哪些操作有强制安全限制——收敛成一组统一原语；其二「自然语言说明书」：开发者用大白话描述设备特性，MHS 据此自动生成一份参考文件；其三「三通道」交互：对话通道、文件通道与结构化通道，智能体可以边试边学、先探索再固化流程。
效果已经可验证。在基因泰克（Genentech）的 BCA 蛋白检测中，接入 MHS 后检测均方根误差（RMSE）较手工流程下降约 76%；在多个实验室自动化场景里，原本数周的集成被压缩到数分钟。发布当日，AWS、优傲（Universal Robots）、斗山、Tecan、丹纳赫、Hugging Face LeRobot、树莓派等约 10 家厂商宣布适配，生态初步就位。Anthropic 还同步放出 1 万个免费 Claude 席位，并在 6 月 Claude Science 的商业闭环上，把「AI + 实验」变成可采购的能力。
但 MHS 仍是研究预览，而非已发布的强制规范，边界必须讲清。物理动作后果不可逆——配错试剂、动错机械臂都可能造成真实损失，因此协议层内置了强制安全限制与人工确认。能力也有上限：在 FrontierChallenge 基准（arXiv:2608.24979）上，即便最优智能体配置，也只有约 1/5（20/97≈20.6%）的科研流程能完整跑通，这反映的是科学智能体普遍的早期瓶颈，而非 MHS 独有缺陷。规范本身也尚未开源。
对开发者与站长而言，MHS 的意义在于把「物理世界的 API」标准化——正如 USB-C 统一了接口。我们已整理出完整的中英对照技术文档，覆盖原语拆解、案例数据、生态对比与同期配套动作，欢迎延伸阅读。
Until now, getting an AI to control a lab device meant each hardware vendor writing its own private SDK and protocol. Swap one microscope or pipetting robot and the integration starts from scratch — automating a workflow could take weeks of glue code. MHS tackles exactly this fragmentation: a common language every device can speak and every agent can understand.
MHS has three layers. The Driver collapses a device's real capabilities — what it can measure, which parameters it can tune, which operations carry hard safety limits — into a set of uniform primitives. The natural-language spec lets developers describe a device in plain words, and MHS auto-generates a reference file. A three-channel interface (conversation, file, structured) lets agents explore first, then freeze a workflow.
The results are verifiable. In Genentech's BCA protein assay, MHS cut measurement RMSE by ~76% versus the manual workflow; in several lab-automation scenarios, integrations that took weeks shrank to minutes. On launch day, ~10 vendors — AWS, Universal Robots, Doosan, Tecan, Danaher, Hugging Face LeRobot, Raspberry Pi and others — announced support, and Anthropic also released 10,000 free Claude seats, extending the June Claude Science commercial loop into a purchasable 'AI + experiment' capability.
But MHS is a research preview, not a ratified standard, and the boundaries matter. Physical actions are irreversible — a wrong reagent or a misplaced robot arm causes real damage — so the protocol bakes in hard safety limits and human confirmation. Capability is also bounded: on the FrontierChallenge benchmark (arXiv:2608.24979), even the best agent config completed only ~1/5 (20/97≈20.6%) of scientific workflows end-to-end, a general early-stage bottleneck for scientific agents rather than a flaw unique to MHS. The spec is also not yet open-sourced.
For developers and site owners, MHS standardizes the 'API to the physical world' — much like USB-C unified connectors. We've compiled a full bilingual technical doc covering primitive breakdowns, case data, ecosystem comparisons and launch-day moves; read on for the deep dive.
打开微信「扫一扫」，分享给好友或朋友圈
关闭