Model Hardware Standard: AI operating physical equipment

Summary

Alek Kemeny, a Member of Technical Staff at Anthropic, opens the presentation by highlighting how integrating AI into laboratories and manufacturing facilities has accelerated scientific experiments from weeks to days. Over the past year of deploying AI models into physical workspaces, teams routinely encountered a major barrier: the lack of a common standard for linking software models to physical machinery. Prior to the Model Hardware Standard, connecting different instruments, such as pairing a camera with a microscope stage, required developers to write bespoke software integrations. As additional devices like pumps, incubators, robotic arms, centrifuges, and spectrometers were added to an experiment, the integration complexity scaled significantly, turning setup into weeks of manual engineering work.

Arco Bast, a Postdoctoral Scientist at the HHMI Janelia Research Campus, explains how the Model Hardware Standard resolves this integration friction. Under the MHS framework, every piece of physical hardware connects once through a single standard interface. Any device that communicates using the standard can interact with any other compatible device, operating at bare-metal speed. An AI agent connects directly to the MHS layer, gaining full access to hardware context and controlling operational parameters across the entire instrument array simultaneously.

The video demonstrates practical implementations of this architecture across multiple organizations. In a collaboration with Danaher, Claude autonomously operates a Leica microscope by adjusting the focus, scanning for bacteria, and deciding which regions of interest to photograph next. At Genentech, a scientist supplied an experimental protocol directly as a PDF document into Claude. The model parsed the instructions, executed the physical experiment autonomously across automated liquid-handling and testing hardware, and resolved unexpected runtime errors overnight without human intervention. In experiments at the HHMI Janelia Research Campus, researchers used Claude and MHS to control real-time microscopy of brain neurons, adjusting positions, capturing deep focal layers, and taking side-view scans dynamically.

Kemeny and Bast conclude that accelerating experimental iteration cycles allows researchers to shift their daily focus away from software configuration toward scientific problem-solving. By removing hardware integration barriers and automating complex testing, the standard aims to accelerate discoveries in material science and biological research, compressing a century of potential progress into a single decade.