Skip to content
AnthropicAnthropic

AI models can now help run physical science experiments

The Model Hardware Standard (MHS) is a new standard for AI agents to safely operate physical equipment in scientific research and advanced manufacturing. MHS began as a collaboration between Anthropic’s Beneficial Deployments team and HHMI Janelia Research Campus. This video tells the story of how MHS was developed and shows how it can be used to accelerate scientific research. MHS is now in research preview with select partners. Read more about how we're learning what AI can do in the physical world: https://www.anthropic.com/news/model-hardware-standard-research-preview

Aug 27, 202611mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

A new standard lets AI control lab instruments safely and fast

  1. The video argues that building and integrating experimental hardware consumes the majority of scientists’ time, slowing down discovery even when the underlying scientific questions are clear.
  2. Anthropic and neuroscientist Arco Bast propose the Model Hardware Standard (MHS), a general method for letting AI models communicate with and control diverse lab devices despite incompatible vendor interfaces.
  3. Demonstrations show Claude operating complex instruments (microscope, robot arm) with safety constraints and iterative guidance, including generating new automation code such as organism tracking with a user interface.
  4. A pharmaceutical lab example illustrates closed-loop experimentation, where Claude adjusts protocol parameters based on observed data (e.g., bubbles affecting liquid handling) to improve outcomes faster.
  5. The broader claim is that AI-to-hardware standards could reshape high-impact domains like drug discovery, quantum computing, and fusion by drastically reducing setup time and increasing experimental throughput.

IDEAS WORTH REMEMBERING

5 ideas

Experiment setup—not theory-building—is the dominant time sink in many labs.

The speakers argue that most scientific time is spent integrating, calibrating, and debugging instruments rather than testing hypotheses, creating a major bottleneck in discovery. AI-driven control aims to shift scientists back toward designing experiments and interpreting results.

Model Hardware Standard (MHS) is a proposed universal bridge between AI and lab instruments.

MHS is presented as a common interface layer that lets an AI model issue high-level goals (e.g., set laser power, move a stage) while the system translates them into device-specific commands. The goal is to make interoperability possible even when each instrument “speaks” a different control language.

Safety can be enforced at the hardware-control layer via bounded, refusal-capable motion constraints.

In the robot arm demo, the system enforces predefined safe boundaries and refuses commands that would exceed them, illustrating how guardrails can be built into the control layer. This suggests AI-driven automation can be paired with hard safety constraints rather than relying only on model judgment.

AI can learn instrument operation quickly but must be treated as an iterative collaborator, not a perfect autopilot.

With a Leica microscope, Claude is shown iteratively exploring settings to acquire a good image, while recognizing risks like crashing a high-magnification objective into a valuable live sample. The workflow is framed as collaborative and iterative because the model can make mistakes and needs feedback.

Rapid prototyping extends beyond scripts: models can generate full experimental tooling (e.g., tracking + UI) in minutes.

Claude is asked to write a tracking program for moving organisms (algae), then improves it by adding a UI so humans can observe and trust what it’s doing. This highlights that automation isn’t only about control—tooling, observability, and usable interfaces matter for adoption.

WORDS WORTH SAVING

5 quotes

That process of building the experiment takes maybe 80% of a scientist's time.

Unknown

The problem is that each device has a different language that it speaks, and getting the devices to talk to each other in their languages is very difficult.

Unknown

What he had built wasn't just applicable to this lab. This idea could be used to have AI run any science experiment in the world.

Unknown

The mere fact that I was able to build this from scratch today, and it achieved it in a matter of minutes, that's insane.

Unknown

This is really the first time in history where we are enabling AI to interact with the physical world in drug discovery. It's definitely historic.

Unknown

Scientific experimentation bottlenecksInstrument interoperability and control languagesModel Hardware Standard (MHS)Safety guardrails and refusalsMicroscope automation and imaging workflowCode generation for tracking and UIClosed-loop optimization in drug discovery

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.