Anthropic is evaluating a new standard for Claude to interface with factory and laboratory equipment.
Anthropic has introduced the Model Hardware Standard in a research preview, which is a specification detailing how hardware functions and how AI can safely manage it, for instance, by controlling the speed and angles of a robotic arm. The European Machinery Regulation will begin to address AI-related safety functions starting January 20, 2027.
The standard developed by Anthropic guides an AI on the operation of machines. The Model Hardware Standard, now in research preview, encapsulates the functions of hardware and its operational guidelines, so that a model does not have to depend on a physical manual or on two knowledgeable individuals.
An example provided by the company involves a robotic arm. A vendor can outline in an MHS file how an AI should maneuver a heavy arm safely, setting restrictions on the allowable speeds and angles.
The underlying concept focuses on accessible knowledge that cannot be utilized. “In many instances, scientific progress is stalled due to the inability to operate the equipment,” remarked Jonah Cool from Anthropic’s life sciences division.
The initiative is rooted in Anthropic's prior accomplishments. “What MCP accomplished for software, MHS will replicate for the hardware sector,” stated Alek Kemeny, referencing the Model Context Protocol that links assistants to applications like Gmail, Calendar, and Slack.
The demonstration of this technology occurred in a laboratory setting. According to Anthropic, a scientist from Genentech sent a PDF of an experimental design to Claude, and the assistant executed the experiment autonomously on hardware equipped with an MHS specification.
The test participants include a broad range of companies. Amazon Web Services, Danaher, Hugging Face, and Raspberry Pi have been engaging with the standard, and a waitlist is now available for others interested.
Among these, two companies have European origins. Raspberry Pi is based in the UK and is publicly traded in London, while Hugging Face was founded by a French team.
The trend in machinery is also leaning toward Europe. NEURA Robotics secured up to $1.4 billion in Germany this year, in a sector of humanoid robotics that TNW believes is more favorable than it appears.
Europe is poised to regulate this specific area. Regulation 2023/1230 will completely replace the Machinery Directive on January 20, 2027, and will address AI-driven safety functions and machinery with self-evolving behaviors for the first time.
This regulatory change will redefine what an MHS file may entail. A specification that limits a robotic arm's speeds and angles contributes to safety and, for high-risk categories, self-declaration of conformity will no longer suffice.
It’s essential to consider the safety record. TNW reported instances where Claude Cowork could breach its virtual machine and access credentials on a Mac, an issue that pertains to software.
Anthropic plans to open-source the framework following the conclusion of the preview. In Europe, those who create that file might find themselves regulating a safety component.
Other articles
OpenAI, Anthropic, Google, and Microsoft advocate for cyber defense to be prioritized at the leadership level.
OpenAI, Anthropic, Google, and Microsoft are advocating for cyber defense to be prioritized at the leadership level. The EU has already implemented the reporting they are seeking.
Zoom features on smartphones are impressive. However, after trying the Vivo X300 Ultra, I realize they may not be suitable for everyone.
I dedicated weeks to photographing with the Vivo X300 Ultra's 400mm ZEISS extender kit. The optics are exceptional, but the weight, setup time, and cost leave much to be desired.
What lies ahead for AI and the steps to achieve it
Frontier models have reached a point of convergence. According to Stanford data, the leading systems now achieve scores that are within 5% of one another. Gediminas Rickevičius, SVP at Oxylabs, suggests that the advantage in competition has transitioned from choosing models to focusing on data infrastructure: specifically, web indexes designed for agents and real-time access layers.
Pollen Robotics’ Microduck is designed to simplify and reduce the costs of training physical AI systems, making the process less delicate.
Pollen Robotics' Microduck is a small biped priced at $399, created to facilitate more accessible physical AI training. It allows developers to test new robotic behaviors on a durable platform that can withstand an occasional fall.
SK Hynix has started construction on a $4 billion memory packaging facility in Indiana.
SK Hynix has begun construction on a $4 billion advanced packaging facility in Indiana. Europe granted state aid for a similar plant but then forfeited it.
Pollen Robotics' Microduck is designed to reduce the fragility and cost of training physical AI significantly.
Pollen Robotics' Microduck is a small biped priced at $399, created to facilitate physical AI training. It allows developers to test new robotic behaviors on a resilient platform that can withstand occasional falls.
Anthropic is evaluating a new standard for Claude to interface with factory and laboratory equipment.
Anthropic's Model Hardware Standard allows vendors to define how an AI can control a machine. Starting in January 2027, Europe will oversee the safety features of AI in machinery.