By teaching language models like GPT-4 chemistry principles and facts, their abilities to perform complex chemical tasks could be greatly improved. Researchers recently concluded a successful experiment in which they trained GPT-4 with chemistry knowledge tailored to the task of designing and executing an experiment. Amazingly, even with a deep understanding of the topic, GPT-4 refused to use this information to synthesize dangerous substances such as heroin or sarin gas.
As powerful as language models like GPT-4 may be, they still require some help to operate at a level of competence that is near that of a qualified chemist. Without additional knowledge and capabilities provided by artificial intelligence tools, experts and complex algorithms, these language models would be insufficient to make substantive progress in the laboratory.
For example, by augmenting the model with additional information about the specific task, researchers have been able to push the boundaries of what GPT-4 can do. With the extra knowledge that the model was given, it became proficient at quickly planning, strategizing and executing sophisticated laboratory tasks. The model even had the capacity to pass initial tests, showing that it is capable of contributing to complex tasks, such as understanding protocols and following them with precision.
The company OpenAI is a research lab based in San Francisco dedicated to discovering and advancing advanced states of artificial intelligence. OpenAI created GPT-4, a deep learning system designed to understand large datasets and generate natural language in the form of text, images and more. The model is a powerful language model based on the “Transformer” technology and has been fine-tuned to such an impressive level that it is able to generate human-level text.
Dario Amodei, Vice President of Research at OpenAI, leads the team in pushing the boundaries of deep learning and artificial intelligence. He has authored multiple papers on AI and is an advisor to several AI startups and organizations. Dario has become a top AI resource for research students and is a long-time believer in the potential of deep learning, the power of open source and collaboration. Currently, Dario and his team at OpenAI are focusing on applying deep learning to personalized medicine. With his leadership, insight and research, GPT-4 is sure to continue to be a powerful tool in the AI toolkit.
