“Autonomous chemical research with large language models,” published in Nature in December 2023 by Daniil Boiko, Robert MacKnight, and Gabe Gomes at Carnegie Mellon, introduced Coscientist, an AI system built around GPT-4 that can carry out chemistry experiments largely on its own.
Coscientist is an agent: the language model is wrapped with tools that let it search the internet and technical documentation, write and run code, and command real laboratory robots through their programming interfaces. Given a high-level goal, it breaks the task down, looks up what it needs, plans a procedure, and executes it on physical hardware. The paper demonstrated this across several tasks, including optimizing palladium-catalyzed cross-coupling reactions, a workhorse of organic chemistry.
The work is significant because it shows a general-purpose language model, not a system trained specifically for chemistry, orchestrating an end-to-end experimental campaign. The intelligence comes from reasoning and tool use rather than from a bespoke chemistry model.
For a general reader, Coscientist and the A-Lab together mark 2023 as the year autonomous experimentation became concrete. They also raise the obvious dual-use question that the authors themselves discuss: the same agent that optimizes a useful reaction could in principle be pointed at a harmful one, which is why safety guardrails are part of the conversation from the start.