“Demis Hassabis believes that training a general model like Gemini on all protein data would be inefficient and negatively impact its other capabilities, such as language skills, due to information efficiency trade-offs.”