OpenAI has launched a new model training approach called Preference Fine-Tuning, allowing users to customize AI responses by comparing desired and undesired outputs. This method enhances the training process for subjective tasks that require tone, style, and creativity. The new framework, Direct Preference Optimization (DPO), simplifies the training of AI models by enabling alignment with user preferences more effectively than traditional Reinforcement Learning from Human Feedback (RLHF). Additionally, the protein language model ProCyon has been introduced, which integrates over 33 million human protein phenotypes through a novel instruction tuning dataset, PROCYON-INSTRUCT. This multimodal foundation model is designed to improve protein phenotype predictions and contextual protein retrieval by leveraging advanced capabilities such as interleaved phenotype-context modeling and multimodal fusion.