Large language models such as GPT-4 have been approaching human-level ability across many expert domains. GPT-4 can accomplish complex tasks in chemistry purely from English instructions, which may transform the future of chemistry.

Large language models (LLMs) predict an output sequence from an input sequence. For example, you could input “ethanol” and it will output CCO — the simplified molecular-input line-entry system (SMILES) representation of ethanol (SMILES are how a chemical structure is written as text). Like any machine learning model, LLMs are fit empirically with a large dataset — generally a large subset of the internet. While they were first pursued for tasks in natural language, such as translating English to French, they can now be used to identify objects in images, predict protein structure and estimate reaction yields, and they are the technology behind the popular ChatGPT.

The recent release of GPT-4 has renewed interest in how LLMs, such as GPT-4, can be applied in chemistry. I have been using the early versions of GPT-4 since 6 months before the release and believe they represent the future of the field — but not by replacing existing computational or experimental methods. Instead, LLMs will transform how we connect our data, computer programs and scientific literature, and how we plan experiments.