The article discusses the application of multimodal large language models (LLMs) in medicine, highlighting their potential to enhance clinical workflows. Multimodal LLMs, such as GPT-4V, can process text, images, and audio, enabling them to assist clinicians in various tasks. These models can transcribe and summarize speech data, generate clinical records, and integrate patient history with imaging data to provide recommendations. They can also recognize text and numbers from images and understand video content for procedural documentation. However, challenges remain, including the risk of hallucinations, where models produce incorrect information, and privacy concerns due to the large amount of data used for training. Additionally, regulating these models is complex, as they require new approaches to testing and mitigation of AI failures. The article suggests that regulators should adapt existing regulations for anticipated applications and use isolated validation datasets to ensure trustworthy validation. Despite these challenges, multimodal AI powered by foundation models shows promise in augmenting the medical workforce in clinical decision-making. The release of GPT-4V is expected to drive future efforts in the responsible development, use, and regulation of multimodal medical AI, improving AI trustworthiness and accessibility in medicine.The article discusses the application of multimodal large language models (LLMs) in medicine, highlighting their potential to enhance clinical workflows. Multimodal LLMs, such as GPT-4V, can process text, images, and audio, enabling them to assist clinicians in various tasks. These models can transcribe and summarize speech data, generate clinical records, and integrate patient history with imaging data to provide recommendations. They can also recognize text and numbers from images and understand video content for procedural documentation. However, challenges remain, including the risk of hallucinations, where models produce incorrect information, and privacy concerns due to the large amount of data used for training. Additionally, regulating these models is complex, as they require new approaches to testing and mitigation of AI failures. The article suggests that regulators should adapt existing regulations for anticipated applications and use isolated validation datasets to ensure trustworthy validation. Despite these challenges, multimodal AI powered by foundation models shows promise in augmenting the medical workforce in clinical decision-making. The release of GPT-4V is expected to drive future efforts in the responsible development, use, and regulation of multimodal medical AI, improving AI trustworthiness and accessibility in medicine.