Gladstone and Stanford University launch an AI hub to help models track experiments as they unfold
Gladstone and Stanford University launch an AI hub to help models track experiments as they unfold
On August 6, Le Cong and Katie Pollard announced a joint AI hub created by Gladstone and Stanford University. A preprint associated with the launch, published on August 4, describes an AI system that maintains a single, continuously updated record of everything important that happens during an experiment.
An AI model might recommend increasing a reagent concentration, and an instrument might then add the reagent to the samples. This step depends on information that often remains in a person's memory or is scattered across separate files: a laboratory technician replaced a reagent lot, an instrument needs calibration, the protocol was revised, or a new result requires another check. Without this information, the model plans its actions from an incomplete account of the laboratory.
In the post announcing the hub, Le Cong described the idea as follows:
“The next generation of laboratories should keep humans in charge while extending their capabilities with AI, connecting scientific intent, machine reasoning, and physical experimentation through scientific world models and an agentic coordination layer.”
The authors of the preprint call this shared memory a laboratory state model. It is intended to contain information about samples, reagents, instruments, protocols, observations, human interventions, and uncertainty. Under this framework, a robot receives an instruction to increase a reagent concentration only after the system has checked the sample volume, the available reagent lot, the instrument calibration, and the protocol constraints. The system then prepares the next step permitted by the protocol or refers the decision to a scientist.
Cong's team previously introduced LabOS, a system that uses glasses and other devices to provide AI with video of a scientist's actions. It compares those actions with the protocol and helps identify errors in real time. The new paper adds instrument data, analysis results, and scientists' decisions to this video record. All of these inputs are intended to update the shared record of the experiment's state.
The authors propose evaluating the system in scenarios that reproduce the course of an experiment. The evaluation would ask whether the system can connect data from different instruments, propose a testable hypothesis, execute a plan reliably, and refer a decision to a person at the appropriate time. The system should distinguish normal experimental progress from a spurious signal in the data, a recoverable failure, or a situation in which the experiment must be repeated, the protocol changed, or the work stopped.
In this architecture, the scientist defines the question, interprets the result, and makes decisions that are unexpected, ambiguous, or high-risk. The AI handles repeatable tasks: it tracks the experimental state, checks protocol constraints, and prepares the data required for the next step.