We made an ontology of actions, an inventory management system, and an experiment scheduler across machines & people. We’ve integrated 40+ scientific instruments – by reverse engineering drivers and using custom hardware adapters.

This is a physical harness for research.

Once given a goal, models can explore inventory and read specs to design experiments as code. The code is converted into equipment control and instructions to people.

Models can then analyze measurements, replan and make a choice about what to try next – just like a person.

But can they do science in the real world?

To answer this we made SciUniverse [Level 1], a scientific benchmark that spans the research process including protocol debugging and bring-up, instrument operation, data analysis, and longer horizon tasks.

We learned models have theoretical scientific knowledge, but miss critical details in physical laboratory work.

They pipette frozen samples, contaminate DNA, fail to account for evaporating solvents, and vortex open containers.

Models doing Software and Mathematics have had the benefit of instant verification loops. Science is an intrinsically physical endeavor.

AI is going to radically change how we do science. We're building facilities where models can execute the entire research process and learn directly from reality.

You can find SciUniverse and our site at:

c5r.net/sciuniverse/