Anthropic 推出新硬件标准 MHS，让 AI Agent 控制物理世界

Anthropic 举例称，Claude 等模型可以调节激光器、通过另一台摄像头检查结果，然后重复这一过程，从而自动校准整套系统。MHS 还可以让 AI 模型对显微镜进行对焦、分析观测结果、判断哪些部分需要进一步观察，再自动将显微镜移动至相关区域以继续实验。

在一段视频中，Anthropic 还展示了 Claude 如何推理并让机械臂抓起一个铝罐，尽管该模型此前并未针对所需步骤接受专门训练。Anthropic 表示，启用 MHS 的模型不必每次都逐步推理，而是可以通过编写 API 脚本来编排多台仪器之间的步骤，并根据实际条件需要进行调整。

视频

Anthropic 在宣传视频中介绍 MHS

Anthropic 表示，MHS 还包含一套标准化标签系统，用于向可能主要在虚拟世界中训练的模型描述硬件在现实世界中的约束条件。其中包括硬件物理特征的编码信息，例如机械臂的重量和活动范围，以及其可调参数、测量选项和强制执行的安全限制。这些标签随后可以整合进参考文件，向没有相关训练经验的 AI 模型快速提供有关设备的关键信息。

目前，Anthropic 表示，正在 MHS 预览期内与“首批科学研究实验室和先进制造商”开展合作，其中包括 Amazon Web Services（Strands Robots）、Hugging Face（LeRobot）、Raspberry Pi、Automata 和 Universal Robots。该公司写道，这些合作伙伴将帮助 Anthropic“建立安全评估体系，并为操作物理设备的 AI 系统制定最佳实践”。在此之后，Anthropic 计划最终将 MHS 发展为一个开源且“与 Agent 无关”的标准，用于整合 AI 与物理系统。

Anthropic 称，在过去一年与科学合作伙伴进行的早期测试中，公司“看到 MHS 缩短了设备集成所需时间，使得在多种实验环境中能够更快地迭代”。

“如果你能更快地检验假设，就能更快地创造通用技术，”Kemeny 在与该公告一同发布的宣传视频中表示。“这就是一个世纪的进步如何能够压缩到十年之内。”

Anthropic gave the example of a model like Claude adjusting a laser, checking the results via a separate camera, then repeating the process to automatically calibrate the whole system. MHS could also allow an AI model to focus a microscope, analyze the results, decide what part needs more observation, then automatically move the microscope to the relevant section to continue the experiment.

In a video, Anthropic also showed Claude reasoning how to get a robotic arm to pick up an aluminum can even though it had not been specifically trained on the required steps. And rather than reasoning through each step each time, Anthropic says MHS-enabled models can sequence steps across instruments by writing API scripts and adjusting them as conditions require.

VIDEO

Anthropic introduces MHS in a promo video

Anthropic says MHS also includes a standardized tagging system to describe hardware’s real-world constraints for models that may have been trained more in the virtual world. That includes encoded information about the hardware’s physical characteristics (e.g., the weight and range of a robot arm) as well as its adjustable parameters, measurement options, and enforced safety limits. These tags can then be integrated into a reference file that can quickly provide an AI model with crucial information about a device it has no previous training experience with.

For now, Anthropic says it is working with “a first group of scientific research labs and advanced manufacturers” during an MHS preview period, including Amazon Web Services (Strands Robots), Hugging Face (LeRobot), Raspberry Pi, Automata, and Universal Robots. These partners will help Anthropic “build safety evaluations and develop best practices for AI systems operating physical equipment,” the company writes. After that, the plan is for MHS to eventually become an open source and “agent agnostic” standard for integrating AI and physical systems.

In early testing with scientific partners over the last year, Anthropic says it “saw MHS reduce the time it took to integrate devices, mak[ing] it possible to iterate faster in a variety of experimental settings.”

“If you can test hypotheses faster, you could create general technologies faster,” Kemeny said in a promo video alongside the announcement. “This is how a century of progress can condense into a decade.”